<%BANNER%>

Acquisition of Novel Perceptual Categories in a Third Language

Permanent Link: http://ufdc.ufl.edu/UFE0041032/00001

Material Information

Title: Acquisition of Novel Perceptual Categories in a Third Language The Role of Metalinguistic Awareness and Feature Generalization
Physical Description: 1 online resource (147 p.)
Language: english
Creator: Gogoi, Divya
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

Subjects / Keywords: acquisition, american, approximant, assimilation, awareness, bengali, bilingual, bilingualism, categories, category, contrasts, cross, discrimination, english, factors, feature, generalization, gogoi, goodness, identification, intelligibility, language, lateral, learning, malayalam, metalingusitic, model, multilingual, multilingualism, native, non, novel, perception, perceptual, retroflex, rhotic, second, spanish, speech, talker, third, training
Linguistics -- Dissertations, Academic -- UF
Genre: Linguistics thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: ACQUISITION OF NOVEL PERCEPTUAL CATEGORIES IN A THIRD LANGUAGE: THE ROLE OF METALINGUISTIC AWARENESS AND FEATURE GENERALIZATION Many factors that contribute towards the perception of second language (L2) sounds and the subsequent establishment of L2 phonetic categories have been the focus of past research. However, learning a third language, unlike SLA, may be influenced by additional factors attributed to the presence of two language systems in a bilingual instead of one language system in a monolingual. Previous third language acquisition research, mainly lexical processing or syntax acquisition, have suggested that being a bilingual provides a positive influence to learners of a third language in attaining general proficiency. Whether bilingualism or a having multilingual benefit as a factor influences the acquisition of novel non-native speech contrasts by adult bilinguals, remains to be explored. Another factor that may facilitate in the phonetic acquisition of a third language is the role that phonetic features may play in the development of new phonetic categories. Features utilized in native contrasts may generalize in the learning of novel non-native contrasts, even if they play a limited or no role in the initial perception of these non-native contrasts. Such factors influencing third language acquisition within specific areas of proficiency like phonological acquisition have not been previously investigated. This study involves a perceptual training experiment examining the influence of factors of multilingual benefit and/or feature generalization. The training incorporates the acquisition of four speech contrasts (dental/alveolar ? retroflex distinction) of a third language by bilingual (Bengali-English and Spanish-English speakers) and monolingual (American English speakers) groups. The perceptual data is gathered through measures like consonant identification, AX discrimination and perceptual assimilation. The perceptual assimilation data is analyzed using the PAM model to determine the extent to which the initial assimilation patterns change towards direction of learning following the training. The results across the testing measures show subtle effects of the multilingual benefit factor playing a role in facilitating the acquisition of the non-native speech contrasts when comparing individual language groups. This effect is observed as robust when both bilingual groups are combined to increase sample size during reanalysis in all tests, except for AX discrimination. The factor of feature generalization does not show an influence on the acquisition of non-native speech contrasts by Bengali-English bilinguals in any of the testing measures. Therefore, the results are suggestive of only the multilingual benefit factor playing a role in enhancing the acquisition process of a third language by adult bilinguals. The perceptual assimilation analysis across language groups show a change of assimilation types after training revealing a shift towards the direction of learning. The bilinguals show a greater shift from SC assimilation type to TC assimilation types than monolinguals. The study thus suggests that multilingualism as a factor facilitates the acquisition of non-native speech contrasts in third language acquisition.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Divya Gogoi.
Thesis: Thesis (Ph.D.)--University of Florida, 2010.
Local: Adviser: Wiltshire, Caroline R.
Local: Co-adviser: Harnsberger, James D.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2011-02-28

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0041032:00001

Permanent Link: http://ufdc.ufl.edu/UFE0041032/00001

Material Information

Title: Acquisition of Novel Perceptual Categories in a Third Language The Role of Metalinguistic Awareness and Feature Generalization
Physical Description: 1 online resource (147 p.)
Language: english
Creator: Gogoi, Divya
Publisher: University of Florida
Place of Publication: Gainesville, Fla.
Publication Date: 2010

Subjects

Subjects / Keywords: acquisition, american, approximant, assimilation, awareness, bengali, bilingual, bilingualism, categories, category, contrasts, cross, discrimination, english, factors, feature, generalization, gogoi, goodness, identification, intelligibility, language, lateral, learning, malayalam, metalingusitic, model, multilingual, multilingualism, native, non, novel, perception, perceptual, retroflex, rhotic, second, spanish, speech, talker, third, training
Linguistics -- Dissertations, Academic -- UF
Genre: Linguistics thesis, Ph.D.
bibliography   ( marcgt )
theses   ( marcgt )
government publication (state, provincial, terriorial, dependent)   ( marcgt )
born-digital   ( sobekcm )
Electronic Thesis or Dissertation

Notes

Abstract: ACQUISITION OF NOVEL PERCEPTUAL CATEGORIES IN A THIRD LANGUAGE: THE ROLE OF METALINGUISTIC AWARENESS AND FEATURE GENERALIZATION Many factors that contribute towards the perception of second language (L2) sounds and the subsequent establishment of L2 phonetic categories have been the focus of past research. However, learning a third language, unlike SLA, may be influenced by additional factors attributed to the presence of two language systems in a bilingual instead of one language system in a monolingual. Previous third language acquisition research, mainly lexical processing or syntax acquisition, have suggested that being a bilingual provides a positive influence to learners of a third language in attaining general proficiency. Whether bilingualism or a having multilingual benefit as a factor influences the acquisition of novel non-native speech contrasts by adult bilinguals, remains to be explored. Another factor that may facilitate in the phonetic acquisition of a third language is the role that phonetic features may play in the development of new phonetic categories. Features utilized in native contrasts may generalize in the learning of novel non-native contrasts, even if they play a limited or no role in the initial perception of these non-native contrasts. Such factors influencing third language acquisition within specific areas of proficiency like phonological acquisition have not been previously investigated. This study involves a perceptual training experiment examining the influence of factors of multilingual benefit and/or feature generalization. The training incorporates the acquisition of four speech contrasts (dental/alveolar ? retroflex distinction) of a third language by bilingual (Bengali-English and Spanish-English speakers) and monolingual (American English speakers) groups. The perceptual data is gathered through measures like consonant identification, AX discrimination and perceptual assimilation. The perceptual assimilation data is analyzed using the PAM model to determine the extent to which the initial assimilation patterns change towards direction of learning following the training. The results across the testing measures show subtle effects of the multilingual benefit factor playing a role in facilitating the acquisition of the non-native speech contrasts when comparing individual language groups. This effect is observed as robust when both bilingual groups are combined to increase sample size during reanalysis in all tests, except for AX discrimination. The factor of feature generalization does not show an influence on the acquisition of non-native speech contrasts by Bengali-English bilinguals in any of the testing measures. Therefore, the results are suggestive of only the multilingual benefit factor playing a role in enhancing the acquisition process of a third language by adult bilinguals. The perceptual assimilation analysis across language groups show a change of assimilation types after training revealing a shift towards the direction of learning. The bilinguals show a greater shift from SC assimilation type to TC assimilation types than monolinguals. The study thus suggests that multilingualism as a factor facilitates the acquisition of non-native speech contrasts in third language acquisition.
General Note: In the series University of Florida Digital Collections.
General Note: Includes vita.
Bibliography: Includes bibliographical references.
Source of Description: Description based on online resource; title from PDF title page.
Source of Description: This bibliographic record is available under the Creative Commons CC0 public domain dedication. The University of Florida Libraries, as creator of this bibliographic record, has waived all rights to it worldwide under copyright law, including all related and neighboring rights, to the extent allowed by law.
Statement of Responsibility: by Divya Gogoi.
Thesis: Thesis (Ph.D.)--University of Florida, 2010.
Local: Adviser: Wiltshire, Caroline R.
Local: Co-adviser: Harnsberger, James D.
Electronic Access: RESTRICTED TO UF STUDENTS, STAFF, FACULTY, AND ON-CAMPUS USE UNTIL 2011-02-28

Record Information

Source Institution: UFRGP
Rights Management: Applicable rights reserved.
Classification: lcc - LD1780 2010
System ID: UFE0041032:00001


This item has the following downloads:


Full Text





ACQUISITION OF NOVEL PERCEPTUAL CATEGORIES IN A THIRD LANGUAGE: THE
ROLE OF METALINGUISTIC AWARENESS AND FEATURE GENERALIZATION




















By

DIVYA VERMA GOGOI


A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL
OF THE UNIVERSITY OF FLORIDA IN PARTIAL FULFILLMENT
OF THE REQUIREMENTS FOR THE DEGREE OF
DOCTOR OF PHILOSOPHY

UNIVERSITY OF FLORIDA

2010
































2010 Divya Verma Gogoi





























To my parents
my husband, Ankur
and our daughter Kaavya









ACKNOWLEDGMENTS

Working on this dissertation has been a challenging and enriching experience for me.

Without the support and encouragement from several people, it wouldn't have been possible for

me to undertake this research. I would first like to thank my academic advisor and committee

chair, Dr. Caroline Wiltshire and my Co-Chair, Dr. James Harnsberger, for their guidance,

constant encouragement and willingness to be available for discussions, their incise comments

and patient feedback, even if I went off track several times have proved invaluable to this work. I

really appreciate Dr. Harnsberger's efforts to pull me out of the statistical 'twister'. I thank my

dissertation committee members Dr. Wayland, Dr. Lord and Dr. Shrivastav for lending their

invaluable guidance as phonetic experts and language specialists towards my dissertation work.

To Dr. Wayland for making international students feel welcomed during our initial stay at UF

and her constant willingness to give feedback whenever sought. I would like to take this

opportunity to thank the Language Learning dissertation Grant Committee for approving the

partial funding for this project. I express my gratitude to all the subjects who participated in this

research. Without their co-operation, I wouldn't have been able to complete the project in time.

To my parents who patiently waited for my visit home, the 'next month', which never

came. They have always supported me to do what I feel is right and then be ready to face the

consequences, good or bad. I thank Bisha, for her immense support in every way during the data

collection process in India. I thank all the wonderful people with whom I had a chance to be

friends with at UF. Especially, Monica, Ashwini, Srikant and my fellow graduate students:

Hanye, Weihua Zhu, Zoe Ziliak, Sharon Barkley and Quinn Hansen, who have been a big

support to me and most importantly I would like to thank my husband, Ankur for being a

constant source of energy that fuelled me to take on a lot of things unbridled.









TABLE OF CONTENTS

page

A C K N O W L E D G M E N T S ..............................................................................................................4

L IS T O F T A B L E S ................................................................................................. ..................... 8

LIST OF FIGURES ............................................. .. .......... ............ ...............9

A B S T R A C T .......................................................................................................... ..................... 1 1

CHAPTER

1 INTRODUCTION ............................................ .. .......... ................................... 13

O v erview .............................................................................................. ......... 13
G o als o f th e stu d y ................................................................................................................. .. 14
L iteratu re R ev iew ................................................................................................................. .. 15
Third Language Acquisition ................... ............................................. 15
Factors Influencing Acquisition of L2/L3 Speech Sounds......................................... 17
P ro p o sed F acto rs ............................................................................................................. 19
M etalinguistic aw areness ......................................... ........................ ............... 19
F feature generalization ................ .............. ............................................. 2 1
P erceptual assim ilation ......................................... .. ........................ ............... 23
Cross Language Speech Perception Models......................................................................25
Speech L earning M odel (SLM )........................................ ....................... ................ 25
Perceptual Assimilation Model (PAM) ......................................................................27
T h e P re se n t S tu d y ................................................................................................................... 3 1
O rg a n iz a tio n ...........................................................................................................................3 4

2 METHODOLOGY .................................. .. .......... ............................. 35

Introduction ................................................. ............................. 35
D description of langu ages .................................................... .............................................. 35
M alay alam .................................................................................................... ....... .. 36
N asals .................................................................................................... ...... .. 3 7
F ric ativ e s .............................................................................................................. 3 9
L ate ra ls ............................................................................................................... 3 7
R h o tic s ...................................................................................................... ........ .. 4 0
B en g ali ........................................................................................................... ....... .. 4 1
S p an ish ........................................................................................................... ........ .. 4 3
American English .............................. .. .......... ............................. 45
S tim u lu s M materials ................................................................................................................ .. 4 6
Stimulus Material Evaluation .................. .............................................................49
Stimuli Evaluation: Preliminary Experiment ............................................................. 50
Stimulus Preparation................................. .. .......... ............................. 51
Subjects ........................................................................................................... 54









B engali-E english B ilinguals ............................................................................ ................ 56
Spanish-E english B ilinguals .................................................................... ................ 57
A m erican E english M onolinguals ................................................................ ................ 58
E x p erim mental D design ....................................................................................................... 59
P perception T raining M ethod ........................................... ......................... ................ 59
P procedure .................................................................................. ....................... 6 1
Fam iliarization .......................................................................................... 62
P retest ............................................................................. ...... ........................... 62
T ra in in g ....................................................................................................................6 5
P o state st ............................................................................................ ..................... 6 6
Task O rder Presentation ........................................................................................66
D ata A analysis ................................................................................... ...................... 67
S u m m a ry ........................................................................................................ ..................... 6 9

3 R E S U L T S ..................................................................................................... ..................... 7 0

Introduction ...................................................... .................. 70
R e su lts .................. ....... ......................... ....................................................... .......... ....... 7 2
Main Effects and Interaction of Independent Variables ................................ ...............72
Examining Evidence for Individual Factors................... ....................................76
M ultilingual benefit ...............................................................................................76
F feature generalization ........................................................................................... 81
R e-A naly sis ....................................................................... ....................... ...................84
Re-analysis: Multilingual Factor .....................................85
An Alternative Re-analysis: Only Place Contrasts.......................................................88
Perceptual Assimilation Results .............................................92
Changes in Assimilation Types after Training.............................................................93
Changes in Assimilation Types: Multilingual Benefit Factor ................................... 95
Summary ....................................................... .................. 99

4 DISCUSSION AND CONCLUSION ..................... ..........................................101

Introduction ............................................... ............... ...................101
O v erv iew o f th e R esu lts........................................................................................................10 1
A ltern ate E x p lan atio n s..........................................................................................................10 4
Test of G eneralization R results .................................................................................105
N necessity for R analysis .............................. ............................................. ............... 107
Caveats in Perceptual Assimilation Analysis ............................................ .................. 111
Results in Relation To Previous Studies............ .............................119
C o n clu sio n ................................................................................................... ..................... 12 5

APPENDIX

A PRELIM IN AR Y SCREEN IN G ........................................... ........................................ 127

B SUBJECT INFORMATION ......................................................................130

C EXPERIMENT INSTRUCTION SHEET............................ ............... ................ 133


6









D PERCEPTUAL ASSIMILATION RESPONSE SHEET ......................... ...................136

E RESULTS IN TABULAR FORMAT .......................................................... 137

L IST O F R E FE R E N C E S ....................................................... ................................................ 139

B IO G R A PH IC A L SK E T C H ................ ............................... ............................................. 147
















































7









LIST OF TABLES

Table page

2-1 Phonem ic consonant inventory of M alayalam .............................................. ................ 37

2-2 Phonem ic consonant inventory of Bengali ................................................... ................ 42

2-3 Phonem ic consonant inventory of Spanish................................................... ................ 44

2-4 Phonemic consonant inventory of American English...................................................45

2-5 Talker information; all talkers are native speakers of Malayalam and born in the state
o f K erala, In d ia ; ................................................................................................................ 4 8

2-6 Result of the preliminary AX discrimination task across three Malayalam speakers.
Mean D prime values averaged across contrasts as well as across vowels ...................51









LIST OF FIGURES


Figure page

2-1 A map of the state of Kerala with cities of origin for the Malayalam speakers. A
smaller map of India on the left hand bottom corner shows the location of the state of
K erala .............................................................................................................. ........ .. 4 9

2-2 An outline map of Mexico, Central America and South America showing the
regional distribution of the subjects .............................................................. ................ 58

2-3 Different trials possible in an AX discrimination task using two tokens of one speech
sou n d s. .......................................................................................................... ...... .. 6 4

3-1 Mean percent correct identification scores averaged over all four contrasts at the
pretest and the posttest level ................................................................... ................ 73

3-2 Mean d' scores of AX discrimination test over all four contrasts at the pretest and
posttest level (M aximum value of a d' score = 4). ....................................... ................ 74

3-3 Mean percent correct identification of contrasts spoken by trained-on talker(posttest)
and new talker (generalization test) ..................................... ...................... ................ 76

3-4 Individual comparison of the monolingual group with BE bilingual group (top chart)
and SE bilingual group (bottom chart).......................................................... ................ 77

3-5 Mean percent correct response in identifying the lateral non-native contrast [ 1-[ ] in
the identification test at the pretest and posttest level................................... ................ 79

3-6 Mean d' scores of discrimination of the lateral non-native contrast [1-[ ] at the pretest
and posttest levels by the subjects in the three language groups. .................................81

3-7 Mean percent identification scores for the nasal non-native contrast [n-nj by the
subject groups at the pretest and the posttest level. ...................................... ................ 82

3-8 Mean percent identification scores for the fricative non-native contrast [S-g] by the
three language groups at the pretest and the posttest level ...........................................83

3-9 Mean percent scores of identification test across all four contrasts by the bilingual
groups (BE and SE) and the monolingual group (AE) at the pretest and the posttest
lev e ls. ...................................................................................................... ........ ....... 8 6

3-10 Mean d' scores of AX discrimination test across all four contrasts, by the bilingual
groups (BE and SE) and the monolingual group (AE) at the pretest and the posttest
le v e ls ....................................................................................................... ........ ....... 8 7









3-11 Mean percent correct response for identification of the three place contrast lateralss,
nasals, fricatives) by bilingual (BE and SE) and monolingual (AE) groups at the
pretest and posttest levels .. ...................................................................... ... ............ 90

3-12 Mean d' scores from the AX discrimination tests for the three place contrasts
(Laterals, nasals, fricatives) by the language groups at the pretest and posttest levels. ....91

3-13 Mean percentage of assimilation elicited from all three language groups, pretest vs
p o state st. .......................................................................................................... ....... .. 9 4

3-14 Percentage of difference scores pretestt to posttest) of assimilation types for the
individual contrasts elicited from all three language groups ........................................95

3-15 Percentages of assimilation types elicited from individual language groups at the
pretest and posttest levels are provided in three different charts............... ................97

3-16 Percentage of difference scores (posttest-pretest) of assimilation types elicited from
BE, SE bilingual groups and AE monolingual group over all four contrasts.................98

4-1 Hypothetical representation of the learning continuum displaying the ambiguity of
ranking the UU, UC and CG assimilation types in order to assess the direction of
le a rn in g ................................................................................................. .................... 1 1 8









Abstract of Dissertation Presented to the Graduate School
of the University of Florida in Partial Fulfillment of the
Requirements for the Degree of Doctor of Philosophy

ACQUISITION OF NOVEL PERCEPTUAL CATEGORIES IN A THIRD LANGUAGE: THE
ROLE OF METALINGUISTIC AWARENESS AND FEATURE GENERALIZATION
By

Divya Verma Gogoi

August 2010

Chair: Caroline R. Wiltshire
Co-chair: James D. Harnsberger
Major: Linguistics

Many factors that contribute towards the perception of second language (L2) sounds and

the subsequent establishment of L2 phonetic categories have been the focus of past research.

However, learning a third language, unlike SLA, may be influenced by additional factors

attributed to the presence of two language systems in a bilingual instead of one language system

in a monolingual. Previous third language acquisition research, mainly lexical processing or

syntax acquisition, have suggested that being a bilingual provides a positive influence to learners

of a third language in attaining general proficiency. Whether bilingualism or a having

multilingual benefit as a factor influences the acquisition of novel non-native speech contrasts by

adult bilinguals, remains to be explored. Another factor that may facilitate in the phonetic

acquisition of a third language is the role that phonetic features may play in the development of

new phonetic categories. Features utilized in native contrasts may generalize in the learning of

novel non-native contrasts, even if they play a limited or no role in the initial perception of these

non-native contrasts. Such factors influencing third language acquisition within specific areas of

proficiency like phonological acquisition have not been previously investigated.

This study involves a perceptual training experiment examining the influence of factors of

multilingual benefit and/or feature generalization. The training incorporates the acquisition of









four speech contrasts (dental/alveolar retroflex distinction) of a third language by bilingual

(Bengali-English and Spanish-English speakers) and monolingual (American English speakers)

groups. The perceptual data is gathered through measures like consonant identification, AX

discrimination and perceptual assimilation. The perceptual assimilation data is analyzed using

the PAM model to determine the extent to which the initial assimilation patterns change towards

direction of learning following the training.

The results across the testing measures show subtle effects of the multilingual benefit

factor playing a role in facilitating the acquisition of the non-native speech contrasts when

comparing individual language groups. This effect is observed as robust when both bilingual

groups are combined to increase sample size during reanalysis in all tests, except for AX

discrimination. The factor of feature generalization does not show an influence on the acquisition

of non-native speech contrasts by Bengali-English bilinguals in any of the testing measures.

Therefore, the results are suggestive of only the multilingual benefit factor playing a role in

enhancing the acquisition process of a third language by adult bilinguals. The perceptual

assimilation analysis across language groups show a change of assimilation types after training

revealing a shift towards the direction of learning. The bilinguals show a greater shift from SC

assimilation type to TC assimilation types than monolinguals. The study thus suggests that

multilingualism as a factor facilitates the acquisition of non-native speech contrasts.









CHAPTER 1
INTRODUCTION

Overview

The globalization of the world has led to many scenarios where people with different

native languages come face to face with each other and the need to communicate arises. The

field of second language acquisition at the cognitive level has a long history of exploration of the

issue of whether bilinguals have a single system of memory representation or separate systems

for each language that they use. Previous studies on this issue have suggested presence of a

single or shared system for more than one language at the lexicon level but acknowledge the

probability of the presence of more language-specific levels as well (see Flege, 1995; Kirsner et

al., 1984; Grosjean, 1992; Krashen, 1981; Kroll and Tokowicz, 2005; De Bot, 1992). SLA

studies have looked into various factors that influence the process of acquiring a new language

by adults who have only one language system. However, the scenario is not that simple.

Acquisition of more than two languages has become more a norm than exception. Therefore,

research in the field of third language acquisition or in broader perspective, multilingualism has

made strides in the last few years.

Since the last decade researchers have turned their attention to investigating the factors that

influence the acquisition of a third language (Cenoz, Hufeisen & Jessner, 2001; Baker, 2001;

Bialystok, 2001; Munoz, 2000; Sanz, 2000). While there has been growing interest in L2

acquisition and bilingualism, many studies have looked into the factors such as age of

acquisition, amount of use of language, context, status of the LI etc. and their effect on learning

L2 either simultaneous or sequential. Fewer studies have looked into third language acquisition

within specific areas of proficiency like phonological acquisition (Enomoto, 1994; Werker,

1986). Previous third language acquisition research have suggested that being bilingual provides









a positive influence to learners of third language in attaining general proficiency (Cenoz &

Valencia, 1994; Ardeo, 2000; Munoz, 2000; Bild & Swain, 1989). These studies were mostly

done on children in classroom settings. Whether or not factors such as metalinguistic awareness

and/or feature generalization enhance the acquisition of the phonetics of a third language for the

bilinguals has not been previously investigated. In the present study, an experiment is set up in

this study to examine such factors after providing training with the new contrast in a laboratory

setting.

Goals of the study

The goal of the present study concerns the exploration of certain factors that may

positively influence the acquisition of a third language by adult bilinguals. Specifically, it

examines the effects of bilingualism in terms of two factors; firstly, benefits of possessing two

linguistic systems often termed as metalinguistic awareness and secondly, transfer of phonetic

feature knowledge from L1/L2 to L3. The study investigates the effects of these two factors on

the acquisition of a third language, Malayalam by two bilingual groups Bengali-English and

Spanish-English speakers and one monolingual group of American English speakers. By

presenting the subjects with novel speech contrasts retroflexx sounds in manners that are lacking

in the subjects' L1/L2 sounds inventory) over a limited period of training, the study aims to

examine if any additive effects of being a bilingual are seen in the course of learning. The

specific sound contrasts from Malayalam that are used as stimuli are lateral [1 1], nasal [n-nj],

fricative [S-g] and rhotic [r--]. The experiment set up for this study includes various tests of

identification, discrimination and perceptual assimilation to observe the effects of the factors in

different levels of perception.









The main hypothesis being tested in the present study is that the bilingual groups are

expected to perform better than the monolingual group in perceiving and identifying the speech

contrasts from the target language. Such results would argue for positive influence of

bilingualism on the third language acquisition in the specific area of phonetics. Within the

bilingual groups, the Bengali-English learner group, who have phonemic experience with

retroflex feature but only in manners other than what is presented in the experiment, is expected

to perform better than the Spanish-English learner group, who have no exposure to retroflex

feature, phonemic or phonetic. By incorporating previous language experience as a factor within

the bilingual groups, the results will provide information for an additive effect of feature

generalization in enhancing the acquisition of speech contrasts from third language over the

effect of bilingualism. The results will also be analyzed in terms of whether they fit the

predictions of the Perceptual Assimilation Model (PAM) proposed by Best & Tyler 2007. The

study aims to examine the changes in assimilation patterns of nonnative speech contrasts by the

listener groups at both the initial perceptual stage and after training stage by adopting the

assimilations types described in PAM.

Literature Review

Third Language Acquisition

A vast corpus of literature exists concerning second language acquisition (SLA) in

adulthood. Research on third language acquisition as a separate entity is a relatively recent

endeavor. Therefore, research in this field is not as extensive. Third language acquisition refers

to the learning of a non-native language by an individual who has already acquired two

languages either simultaneously (before puberty) or sequentially. The scenario of third language

acquisition although quite similar to SLA still presents differences since, unlike second language

learners, third language learners have more experience as language learners and have access to









two linguistic systems when acquiring a third language (see Cenoz, 2003). Most of the previous

studies on third language acquisition have looked into the question of whether bilingualism

facilitate the acquisition of third language? (Cenoz, Hufeisen & Jessner, 2001; Baker, 2001;

Bialystok, 2001; Munoz, 2000; Sanz, 2000) Most of these studies have been conducted on school

children in a classroom setting. The focus of these studies have been primarily to examine the

level of general proficiency in the target language being learnt through various linguistic tasks

like word awareness, grammar, speaking, reading and writing. Their results have suggested that

being a bilingual provides positive influence on the acquisition of the third language (Cenoz &

Valencia, 1994; Ardeo, 2000; Munoz, 2000; Bild & Swain, 1989). In addition to the effect of

bilingualism, third language acquisition also presents a diverse range of other factors that may

influence the learning of the target language. These factors are quite similar to the typical factors

that affect acquisition of the second language, such as age of acquisition or the amount of use of

the language (see e.g. Flege, 1998; Flege et al., 2002) or phonetic, phonemic or acoustic factors

(Polka, 1991), coupled with other influential factors that are traditionally explored in the field of

bilingualism such as L1/L2 status, socioeconomic factors, degree of bilingualism etc (see

Hammarberg, 2001). More specific to the current study, the effect of bilingualism, also termed as

metalinguistic awareness, along with feature generalization, here referred as the transfer of

phonetic/phonemic feature knowledge from L1/L2 sound inventory to facilitate in the acquisition

of L3 sounds, is examined in the area of phonological acquisition of third language. The

following section provides a background review of certain factors that may influence acquisition

of sounds in both L2 and L3 along with the detailed review of the proposed factors -

metalinguistic awareness and feature generalization.









Factors Influencing Acquisition of L2/L3 Speech Sounds

Through extensive research many factors that contribute towards the perception of second

language (L2) sounds and subsequent establishment of L2 phonetic categories have been

examined. Based on such factors, theories on cross language speech perception by naive as well

as experienced listeners have been developed (e.g., perceptual assimilation model, or PAM

(Best, 1994, 2007; Flege, 1995). The term naive listener here refers to a listener who has no prior

exposure to the non-native sounds of the target language and an experienced learner is

acquainted with the non-native sounds although he/she has not formed any phonetic categories

corresponding to these sounds. Phonetic, phonemic and acoustic factors, among others, have

been determined to influence the learning of L2 sounds (Polka, 1991). Previous studies that have

addressed such factors are briefly discussed below.

The factor of phonetic experience defined as, "experience with specific phonetic segments

as allophones or free variants of native phoneme categories" (Polka, 1991: 2962), has been

explored in cross-language speech perception research. It has been hypothesized (Werker et al,

1981; Tees and Werker, 1984; Strange and Dittmann, 1984; Polka, 1992) that a non-native

contrast that contains a phonetic segment identical to a segment frequently experienced only in

phonetic contexts by a subject in his languagess, is more easily differentiated by the learner as

opposed to non-native contrasts which may not employ allophonic experience. The findings from

Werker et al (1981) and Tees & Werker (1984) showed naive American English adults'

increased difficulty in differentiating (after training) the [t ] place contrast compared to the

[dh-th] voicing contrast from Hindi. They credited allophonic experience as an explanation for

these results suggesting that experience with VOT contrasts in the subjects' native language

(AE) facilitated the differentiation of the non-native voicing contrast. However, a limited number









of subjects who had experience with the Hindi language in their early childhood were able to

perceive the place contrast after training, suggesting early linguistic experience facilitates better

perception of the place contrast than the AE subjects with no early life experience with Hindi

(Tees & Werker, 1984). Polka's (1992) study tested the above given hypothesis but found no

conclusive results due to presentation order effects. However, the allophonic experience

hypothesis was contested in subsequent studies (Best et al,. 1988; Pruitt et al,. 2006), suggesting

that phonetic experience alone could not explain the variability found in perceptual performance

on different non-native contrasts and could even have a detrimental effect in the perception of

non-native contrasts. Phonemic and phonetic experience together (presence of linguistic

experience), are able to better explain the facilitation effects that are seen in the perception of

non-native contrasts (Best et al,. 1988; Sundara et al,. 2006).

Another factor of acoustic salience that may influence the acquisition of L2/L3 sounds was

proposed by Burnham (1986) to account for variable perception levels of non-native contrasts.

Burnham claimed that the distinction between the easily discriminable and the poorly

discriminable non-native speech contrasts can be explained in terms of their inherent

psychoacoustic basis. He used the terms "robust" and "fragile" to maintain the psychoacoustic

distinction between non-native contrasts. According to him, a contrast was robust if the reliable

acoustic cues in that speech contrast are too salient to be lost during memory demanding tasks.

However, psychoacoustic salience individually as a factor can explain perceptual performance

only when some independent acoustic cue measures can be determined. Further research needs to

be conducted in this respect.

The above discussed factors have been determined to substantially influence the perception

of non-native speech sounds of a target language. However, without explicit definition, a target









language can be assumed to be a second language (L2), a third language (L3) or a novel

language (language unknown to the listener). The area of third language acquisition (or

multilingualism) until recently was considered equivalent to second language acquisition (SLA)

(as noted in Hammarberg, 2001). However, learning a third language, unlike SLA, may be

influenced by additional factors like effects of bilingualism (Hoffman, 2001; Sanz, 2000;

Chenoz, 2001; Bialystok, 1992, 2001). The section that follows provides detailed information on

the factors explored and presents them in the context of the current study.

Proposed Factors

Metalinguistic awareness

Metalinguistic awareness allows the individual to step back from the comprehension of an

utterance; to attend to the structural features of the sounds in order to abstract the knowledge of

its distinguishing cues in a speech contrast, a task which demands certain cognitive and linguistic

skills (Malakoff, 1992: 518). This awareness helps a bilingual to acquire non-native contrasts

more easily than monolinguals. The assumptions in the formulation above are based on the

results of studies on lexical processing and word learning by bilingual and monolingual children

showing that bilingual children are more adept at doing metalinguistic tasks and have cognitive

benefits from knowing two or more languages (Peal and Lambert 1962; Munoz, 2000; Klein,

1995; Sanz, 2000 etc.). The study by Klein (1995) showed that during lexical acquisition,

multilinguals learned the same number of lexical items as monolinguals but at a quicker rate of

acquisition. Klein based the explanation for this phenomenon on enhanced cognitive skills in

multilinguals which helped them tease out the potential relevant data for resetting the particular

parameter, preposition stranding in this case, for the new language. Numerous studies looking at









cognitive processing1 at a linguistic level have explored the concept of metalinguistic awareness

in bilingual and multilingual children as well as adults (Bialystok, 2004). The studies emphasize

the fact that bilinguals with near native proficiency in both languages would show evidence of

metalinguistic awareness (Bialystok 1992, 2001; see also review in Jessner, 2006). In her study,

Sanz (2000) suggests that metalinguistic awareness may be an enhancing factor in language

processing. She promotes bi-literacy as the cause of heightened metalinguistic awareness.

"... metalinguistic awareness, which results from exposure to literacy in two languages, gives

bilinguals the capacity to focus on form and pay attention to the relevant features in the input"

(Sanz, 2000: 36).

This study extends the concept of metalinguistic ability from the field of lexical processing

to the field of phonetics. This core concept of metalinguistic awareness has not yet been explored

in the field of cross-language speech perception and in the area of phonological acquisition of a

third language. It is yet to be seen if implicit knowledge of two languages, in the case of phonetic

features in the sound systems, can result in heightened capacity to acquire non-native speech

sounds. However, the concept of metalinguistic awareness is a very broad usage for various

metalinguistic skills. This study specifically assesses metalinguistic awareness/ metalinguistic

ability from the perspective of the effects of bilingualism where having two or more languages,

well established within the language system of an individual benefit him for linguistic processing

over a monolingual individual. Therefore, rather than using the term metalinguistic awareness

loosely, the more operational term of multilingual benefit factor will be used in this study.

According to previous studies in lexical processing, the bilinguals may function at a faster rate

and at a higher level for reorganizing information in order to deal with the high-demand task of

'For this study the term cognitive processing is operationally defined as the act of abstraction of relevant information
from a speech utterance, required for a linguistic skill, as a result of higher level of concept formation.









acquiring new non-native speech contrasts when compared with monolinguals. The assumption

that bilinguals develop meta-linguistic awareness which benefits them over monolinguals in

acquiring a third language, can be examined through the experimental design set in which

bilinguals are compared to monolinguals in the acquisition of speech sounds in the third

language, as is done in this study. For instance, in the current study scenario, the retroflex

feature, which is non-existent in the Spanish-English bilingual's phonological space, may still be

acquired by them at a relatively faster rate of acquisition than English monolinguals. This may

reflect that bilinguals employ their metalingual knowledge to pay attention to distinctions in

auditory-acoustic stimuli and to tease apart relevant acoustic cues in order to accurately

distinguish between the newly acquired speech contrasts. The results of perceptual performance

after training would be compared to results of the monolingual group. Therefore, in case of

successful and faster acquisition, results would suggest the presence of meta-linguistic awareness

among bilinguals (Spanish-English) which facilitates the acquisition of a novel contrast without

the limitation of feature generalization of an existing feature.

Feature generalization

Features, an important set of information contained in the established phonetic category,

play a crucial role in the development of new phonetic categories (Nosofsky, 1986,1987;

Kruschke, 1992; Lively et al., 1993; Francis and Nusbaum, 2002). Features may refer to

phonemic attributes of the sound system of a language, phonetic features or acoustic cues

attached to the specific segments of the system. According to Francis and Nusbaum (2002)

features can be stretched or shrunk according to relative weights given to them. For this study,

features refer to the phonetic features within a phonemic segment of a language. The factor of

feature generalization or feature productivity as defined by Polka (1992: 40) is "whether

experience with a particular place distinction facilitates perception of the same phonemic place









distinction in a different manner context..." The factor of feature generalization, operationally

defined for this study, refers to the concept where the listener being sensitive to individual

features in his/her native phonetic categories is able to identify the same feature in a novel non-

native contrast and thus develop new phonetic categories. Few studies have looked at the feature

generalization phenomenon in the field of cross-language speech perception. Polka (1992) and

Harnsberger (1998) both found results contrary to the expected results in this study that feature

generalization will reveal better perceptual performance. However, one factor that could explain

their results is that these previous studies involved only the perceptual presentation of new non-

native contrast to nalve listeners, rather than conducting a perceptual training experiment and

subsequently observing any effects of feature generalization during a perceptual testing phase.

The present study is different from previous works in that it takes the next logical step and

examines the role of feature generalization in the perceptual performance data (posttest data) of

TRAINED listeners as opposed to nalve listeners (see Polka, 1992; Harnsberger, 1998).

Recently, a study by McAllister, Flege, & Piske (2002) found that experience with the duration

feature did facilitate native American English speakers' perception of Swedish vowels

(phonemic lengthening) more than that of native Spanish speakers (both subject groups had lived

in Sweden for at least 10 years). The difference between McAllister et al.'s study and others was

that the subjects were experienced learners of Swedish and used the target language often. This

provides support for the hypothesis of feature generalization. To test for feature generalization in

the current study, a group of Bengali-English bilinguals (BE) undergo perceptual training to

acquire the speech contrasts of retroflex nasal as opposed to alveolar nasal. The Bengali

language has extensive use of the retroflex feature for stops in its phonemic inventory but lacks

this feature in manners like nasal stop [rl], lateral approximant [1,] fricative [f] and central









approximant [t ]. An observation of faster acquisition of these novel sounds by the BE bilingual

group when compared with either the bilingual Spanish-English group or the monolingual

American English group, will indicate a successful generalization of the feature retroflex from

stops to novel manners of articulation like nasal stops. The results of training on the identical

stimuli from the group of English monolinguals would be observed as a control group.

One criticism on the choice of subject group could be that the phenomenon of feature

generalization could be evident in monolingual Bengali speakers as well. The current study deals

with feature generalization as a factor of enhancing learnability of non-native speech contrasts,

regardless of whether the learner is a bilingual or monolingual. However, taking bilingual

learners of L3 for this study provides us with the opportunity to assess both feature

generalization and multilingual benefit within a single group of subjects. Moreover, if the results

do not provide evidence for multilingual benefit (in the Spanish-English (SE) group) but show

enhanced learnability in the Bengali-English group, it would suggest that it is not the factor of

multilingual benefit that is facilitating learnability in Bengali-English but the factor of feature

generalization. On the contrary, if the multilingual benefit factor is observed in the SE group,

then the feature generalization factor is assumed to have an additive effect that can be observed

in BE group as significantly higher performance level than SE group.

Perceptual assimilation

The third focus of the research project concerns the role of the bilinguals' particular

perceptual category inventory in the acquisition of non-native contrasts from a third language.

Looking at the perceptual assimilation patterns elicited from the learners at the post-training

stage will provide an opportunity to examine the formation of new perceptual categories, if any,

developed after a limited time of training. Moreover, perceptual assimilation analysis would be









another dependent measure to test for performance level of acquisition among the subject groups.

This measure will show greater ecological validity as opposed to just perception tests conducted

in previous studies that looked at the acquisition of L2/L3 through perceptual assimilation.

Unlike a perception test, training involves feedback on performance, an essential aspect of

learning. Since a training experiment where an individual is perceptually trained on non-native

speech contrasts would simulate more closely the natural learning process, eliciting and

examining perceptual assimilation (PA) types at the post-training stage and comparing them with

pre-training PA types will prove to be more ecologically valid than simply eliciting the PA types

in a perception experiment without any controlled training.

Previously, several assimilation analytical models have been developed to account for the

influence of native perceptual categories on the perception, production, and acquisition of non-

native speech sounds, including the Speech Learning Model (SLM) and, of greater interest for

this project, the Perceptual Assimilation Model (PAM) since it predicts the assimilation patterns

not only at the initial exposure stage but also predicts learnability beyond the first stage.

Likewise, the current study also looks at the assimilation patterns at the post-training stage. PAM

concerns the discriminability and learnability of non-native contrasts based on the relationship

between the non-native and native category inventories. These relationships have been

encapsulated in several assimilation types: Two-category (a non-native contrast that is highly

similar to a native contrast), Single category (a non-native contrast that is consistently identified

as a good exemplar of a single native speech sound), Category goodness (a non-native contrast

consistently of two speech sounds that are consistently identified with a single native sound but

differ in goodness of fit), uncategorizable, either UC or UU (involving one/both non-native

speech sounds that are not consistently identified with a particular native category) and non-









assimilable (a contrast that is perceived in a nonlinguistic manner). These assimilation types have

been discussed in detail in the next section. The present study tests these predictions of the PAM

model in two ways: 1) Determining the extent to which initial patterns of assimilation change, if

at all, following extensive training experience and 2) measuring the extent of assimilation pattern

change in individual subject groups in order to assess any effects of the influencing factors

explored in this study.

The following section discusses the above mentioned cross language speech perception

models that have been developed predicting perceptual assimilation patterns of naive listeners as

well as subsequent changes in assimilation patterns in the learning of a second language by

experienced learners.

Cross Language Speech Perception Models

Since the advent of speech perception studies on non-native contrasts during the 1980-90s,

various speech perception and acquisition models that address bilingualism have been proposed.

This section of the chapter describes the pertinent models and examines the core claims of each

model.

Speech Learning Model (SLM)

The speech learning model focuses on the L2 phonological acquisition in production and

perception (Flege 1995). It is claimed in this model that many second language production errors

have an underlying perceptual basis. The model considers the factor of perception as the

underlying principle behind the acquisition of L2 speech sounds and effective learning. The

model propagates the notion that if learners of second language sounds are not able to accurately

perceive particular L2 sounds, the consequent production of those L2 sounds will be inaccurate.

One of the preliminary assumptions of the model is that the production and perception of

phonetic systems "remain adaptive over the life span" and they are susceptible to reorganization









to accommodate L2 sounds, when encountered during the life span. Another postulate of this

model is that "bilinguals strive to maintain contrast between LI and L2 phonetic categories (long

term memory representations) which exist in a common phonological space" (Flege, 1995: 239),

bringing forth the relevance of the phonetic /phonemic contrasts between the LI and L2 sounds,

in the process of acquiring the second language. Hypotheses of this model reflect the need to

acknowledge the phonetic differences between the LI and L2 sounds that may be phonetically

closer or more distant, which may influence the perception of L2 sounds by the learners.

The factor of mental representation or categorization of LI and L2 sounds has been

directly addressed in this model. According to SLM, if the bilinguals learn to distinguish some

phonetic differences between the L2 sound and the closest LI sound to it, a new phonetic

category is likely to be established for that L2 sound. Nevertheless, the category formation for an

L2 sound can be blocked if the mechanism of "equivalence classification" comes into play. In

other words, for the bilingual the LI and L2 sounds can be perceptually linked and therefore, one

phonetic category is used for both sounds, which may eventually resemble each other in

production. The term implies that 'equivalent' or similar sounds are difficult to acquire due to the

perception of these specific LI and L2 sounds by the listener as being quite similar (non-

perceivable phonetic distance by listener), whereas 'new' or dissimilar (terms used

interchangeably) sounds are easier because the listener easily distinguishes the phonetic

differences.

SLM also predicts that if a new phonetic category for an L2 sound is in fact established,

the phonetic system would try to maintain contrast between the existing LI phonetic category

and the newly formed L2 category since both LI and L2 categories are assumed to 'exist in a

common phonological space' (Flege, 1995: 239)). Thus, the new L2 category established may









not be identical to the native speaker category, that is, the target sound. Therefore, the

predictions for mental representations of the LI as well as L2 sounds are well documented in this

model.

The model differs from an older contrastive analysis approach as it proposes a less abstract

phonetic level of analysis than phonemic level of analysis. Contrastive analysis of two languages

involves the examination of phonemic differences between the inventories of those languages.

According to one SLM hypothesis, learners perceptually relate positional allophones in the L2 to

the closest positionally defined allophone in the LI. On the other hand, LI for L2 substitution

described in contrastive analysis is seen as either failure to discern the phonetic differences

between the LI and L2 sounds or inability to articulate the correctly perceived sound.

According to SLM, during the course of second language acquisition, the phonetic

category that has already been established for an LI sound during childhood evolves gradually if

it is perceptually linked to an L2 sound. Such developments are influenced by age of learning

and perceived phonetic distance between an LI sound and the closest L2 sound. The model

claims that the greater the perceived cross-language phonetic distance, the more probability for

the development of a separate phonetic category for the L2 sound. Also, age of learning (AOL)

determines how much perceived phonetic distance is required to stimulate the process of separate

category formation for the L2 sound (early AOL + smaller perceived phonetic distance needed =

category formation). In summary, the model claims that the most L2 production errors observed

have an underlying perceptual problem.

Perceptual Assimilation Model (PAM)

The Perceptual Assimilation Model (Best, 1995, 2007) lays its foundation on the

phonetic/phonological nature of the L2 speech sounds versus LI sounds, similarly to the other

speech perception models. The preliminary assumption of the model is that the nature of the









discrepancies and similarities that a non-native listener picks up in the auditory signal may be

perceived in terms of articulatory or acoustic properties. The model takes the 'ecological

perspective' and relates to the Direct Realist Theory (e.g., Gibson, 1991; Gibson, 1966, 1979)

that the listener's perception is stimulated mostly through the evidence about the articulatory

simple gestural events in the speech signal (Best, 1995). The model extends the direct realist

theory in the way that the adult listener perceives information about gestural similarities in the

L2 speech with the native phonemes, when receiving the non-native signal. According to the

model, there is no two-extremes phenomenon (full assimilation or none) in perceiving non-native

speech. Also, all non-native contrasts do not pose same level of difficulty to the listener. The

perceptual difficulty should be predictable from the differences in the gestural

similarity/discrepancy between non-native and native sound distinctions. There are five main

possibilities during perception that the model predicts as outcome of listeners' mental

representation of non-native speech contrasts. The predictions of the discrimination task are also

given alongside the category type speech contrasts.

Firstly, non-native speech contrasts may be treated as two category type (TC). In this case,

the speech contrasts may be gesturally similar to two different native phoneme categories

making the non-native speech contrast easily discriminable by the LI listener. An adult listener

is expected to show near native-like discrimination of this type of non-native speech contrast.

Secondly, non-native speech contrasts may be treated as single category type (SC) if

listener's perception of the non-native speech contrast is equally good or equally poor and both

sounds are assimilated to a single category.

Thirdly, non-native speech sounds may show differences in category goodness type (CG).

When the non-native contrast is assimilated to a single category but one member of the contrast









is seen as a better token of the native phoneme and the other as a marginal token of that category,

this makes the discrimination easier than the SC type but not equivalent to TC type of

assimilation. Again, because category goodness type of pairs may be perceived with varying

degree of discrepancy between the two phones, the level of discrimination may vary.

Finally, Uncategorizable type (UU) are formed where even though none of the phones

from the non-native pairs is assimilated with any native phoneme category, yet the phones are

considered as speech sounds of a foreign language. For Uncategorized versus categorized type

(UC), where only one phone of the non-native pair is assimilated to a native phoneme category,

the adult listener is expected to show a very good discrimination performance for this type of

non-native speech contrast.

Therefore, concluding from the above given possible outcomes of discrimination

performance, the pattern for an adult listener would be highest discrimination for Two category

types, moderate to good discrimination for both category goodness type, non assimilable type

(for different reasons) as well as for the uncategorized versus categorized type, and lowest for

the single category type. The uncategorizable type varies from good to poor depending upon the

acoustic proximity of the two phones in the pair. Also, PAM assumes a strong LI phonological

system influence on the perception ability. The PAM model, in conclusion, emphasizes that the

perception of non-native segments is performed according to their similarities or discrepancies

from the native segmental system, being in close proximity to the native phonological space as

well as with each other (within the contrast) (Best, 1995).

Although the model provides intricate details of perceptual abilities of adults for non-

native speech contrasts with highly probable outcomes for the categorization of the non-native

phones by the listener, the model does not really deal with the learning process of these non-









native phones by the LI speaker, unlike SLM. The earlier PAM model provides the predictions

of perception of non-native speech contrasts by listeners who have not had any experience with

L2. These discrimination outcomes provided in the model are based on the first exposure of the

listener to these sounds. The learning process that takes place after the learner is exposed

repeatedly to the L2 sounds and the reorganization of the LI phonetic categories that takes place

due to the influence of the L2 phones is not addressed in this model.

These differences are addressed in the revised PAM model (Best & Tyler, 2007). The

Perceptual Assimilation Model extends its predictions from the earlier model focusing on the

nalve non-native listener to a revised version of the model where it accounts for the experienced

L2 learner apart from the initial encounter with L2 as a nalve listener. Best refers to the revised

version as PAM-L2. Four possible cases of L2 minimal contrasts are discussed starting with

PAM's initial predictions.

Only one L2 phonological category is perceived as equivalent (perceptually assimilated)
to a given LI phonological category (TC). The discrimination of the speech contrasts
would be excellent as the LI learner would always perceive the fully assimilated L2
phone separate from other less assimilated L2 phones. The L2 learning for this L2
phonological category will not take place as it has been entirely assimilated and
recognized with an LI phonological category.

* Both L2 phonological categories are perceived as equivalent to the same LI phonological
category, but one is perceived as being more deviant that the other (CG). The
discrimination level for this type would range from moderate to good. During the
learning process, the learner is likely to gradually form a new phonological category for
the perceived poor/deviant phone of the LI phonological category. The good exemplar
L2 phone may remain assimilated to the LI phonological category.

* Both L2 phonological categories are perceived as equivalent to the same LI phonological
category, but as equally good or poor instances of that category (SC). The listener will
poorly discriminate the contrast since the two phones would be perceived as having very
little phonetic difference and be considered as exemplars (either equally good or bad) of
single LI category. For the learners to distinguish these two phones apart,relevant
phonetic differences between the phones will have to be learnt, forming a separate
phonological category over time.









* No L1-L2 phonological assimilation: There could be two situations with these types of
minimal contrasts. The listener may not perceptually assimilate either of the phones with
any LI phonological category. If the two L2 phones are in existence with relatively larger
distance in the learner's LI phonological space, the learning process will be easier for the
listener. However, if the L2 phones are relatively close to each other in the LI
phonological space, the listener would be not able to discriminate them easily.

The PAM-L2 model makes the assumption that the learner begins the acquisition of L2 as

a naive but phonologically sophisticated LI listener. With the initial introduction of the phones

of L2 and new L2 phones perceptually assimilated with LI phonological/phonetic categories, the

learning process for the L2 learner starts. The PAM-L2 model makes predictions for the learning

process of this, now experienced, L2 learner. The target phonological system, that the learner

may achieve over a life span, which the model hypothesizes, is that of a common LI and L2

system whereby phonetic categories from both LI and L2 share a common phonological space.

All these models of cross-language speech perception look at the second language learning

by adults. These models can be extended to look at the factors facilitating third language

learning. The current study proposes factors like metalingual benefit and feature generalization

as some of those factors that may facilitate third and/or second language acquisition.

The Present Study

This study is an attempt to provide a window to phenomena not explored well in the field

of cross-language speech perception and phonetics. Firstly, the multilingual benefit concept has

been attested only in psycholinguistics, lexical and cognitive processing or word learning fields.

Moreover, until recently, multilingual benefit was examined only in bilingual children. The

current study will test this factor's influence on third language acquisition in the area of cross-

language speech perception and third language learning. Secondly, the phenomenon of feature

generalization has been evaluated in studies only dealing with naive listener's perceptual

performance. The main experiment of the study deals with training the subjects and testing their









perceptual performance in order to seek evidence for feature generalization. Observing the

perceptual performance of a trained listener rather than a naive listener may be the missing factor

in finding the role of feature generalization in facilitating the learnability of non-native phonetic

categories. Apart from the pioneering steps in exploring metalinguistic ability and feature

generalization, the study adopts the assimilation types established by the revised PAM model and

tests the predictions laid out by addressing the issue of the relation of initial perception

modifying into different assimilation patterns during the course of training. The relevance of

chosen stimulus set and the appropriate subject groups is discussed as follows.

A stimulus set which is relatively difficult to discriminate based on acoustic differences is

archetypal for this kind of training study so as to bring out subtle distinctions in the perceptual

performance of the various subject groups. Dental/alveolar-retroflex place distinction as a speech

contrast is considered to have a reasonable level of difficulty in discrimination due to differences

in spectral features as opposed to temporal features in speech contrasts (Pruitt et. al., 2006;

Polka, 1991; Strange and Dittmann, 1984). Therefore, dental/alveolar-retroflex place distinction

was chosen as the place contrast for the non-native speech contrast as the stimulus set. The

speech contrasts were taken from Malayalam, recorded as natural stimuli from eight speakers.

Malayalam was chosen as the target language as it has a rich inventory of consonants employing

the retroflex feature extensively in various manners. This allows a flexible choice of speech

contrasts as per the requirements of the study. An appropriate and diverse range of challenging

non-native speech contrasts were selected from Malayalam for the study.

1) [ 1 [ ] (alveolar versus retroflex lateral approximant)
2) [n r. ] (dental/alveolar versus retroflex nasal)
3) [J- ] (palato-alveolar versus retroflex fricative)
4) [ r -t ] (alveolar tap versus retroflex approximant)









These speech contrasts differ primarily in place of articulation, although the [ r 4 ]

contrast also differs in manner. The dental/alveolar-retroflex place distinction, common to all of

the chosen speech contrasts, allowed examination of factors (a) feature generalization and (b)

multilingual benefit in three populations of subjects. Furthermore, selection of four different

manners with same place contrast allows for the examination of the learnability level of these

speech contrasts. They can also provide new data testing the predictions of PAM in perceptual

assimilation.

Three subject groups were chosen for this study. Firstly, the reason for choosing these

particular speech contrasts, when looking at Bengali-English (BE) language group, is primarily

for the purpose of evaluating feature generalization. The speech sounds with retroflex feature [ [,

r, -4 ] are not found in Bengali. However, Bengali employs the retroflex feature quite

productively but only in the oral stop and flap manners of articulation. Therefore, if the retroflex

feature gets generalized from existing phonetic categories to newly formed phonetic categories,

the evidence of any variability in generalization of the feature in different manners will be

reflected during the acquisition of these four new speech contrasts.

Secondly, Spanish speakers altogether lack any experience with the dental/alveolar-

retroflex place distinction in phonemic or phonetic environments. However, as the participants of

this study are bilingual (Spanish-English), they are assumed to have phonetic experience similar

to American English speakers, who have alveolar place distinction phonemically and sometimes

have allophonic experience with retroflex oral stops (See Polka, 1991). No prior native

languages) experience with the retroflex feature makes the Spanish-English bilingual group

ideal to examine multilingual benefit factor. This bilingual group, having significant









improvement in identification of stimulus sets after training, if any, can be safely attributed to

multilingual benefit.

The third population of subjects chosen for this study is a control group. Monolingual

American English (AE) speakers were chosen in order to control for the bilingualism factor.

Also, American English speakers lack dental/alveolar-retroflex place distinction at the phonemic

level. However, retroflexion of alveolar stops in particular environmental contexts (e.g. [ 4jai]

'dry' (Polka, 1991)) is sometimes found in the speech of American English speakers.

Organization

This chapter provided an introduction to the current study along with a review of previous

literature regarding cross-language speech perception as well as the posited factors. A detailed

discussion of the research questions evaluated by the study along with the relevance of the

stimuli and subject populations were presented. Chapter 2 gives a comprehensive description of

the phonemic and phonetic inventory of relevant segments in the four languages utilized in the

experiment. It also provides the experimental design of the perceptual training and testing

paradigm used for the main experiment of the study. Chapter 3 presents experimental findings of

the training experiment undertaken in this study. Lastly, Chapter 4 summarizes the findings and

discusses the implications and limitations of the current study.









CHAPTER 2
METHODOLOGY

Introduction

The experimental design of this study is directed towards evaluating the hypotheses that (a)

feature generalization and/or multilingual benefit may facilitate the learnability of non-native

speech contrasts by bilinguals and (b) assimilation patterns in the initial perception of non-native

speech contrasts can be used to predict subsequent learnability in perception and accuracy in

production. The study involves the perception training experiment which consists of learning of

novel non-native speech contrasts. The next section describes the phonemic segments and their

phonetic realizations in the four languages employed in the main experiment. The following

sections describe the preparation and preliminary evaluation of the stimuli for the main

experiment and discuss the recruitment of various language groups as subjects for the current

study. It also highlights the appropriateness of the choice of the three language groups in order to

directly examine the research questions put forth in the study. Relevant literature review on the

method of training and their underlying theoretical assumptions is also provided in subsequent

section. Towards the end of this chapter, the adapted experimental design from the perception

training paradigm is described in detail.

Description of languages

The experiment outlined in this chapter entails the involvement of four languages. The

perception training stimuli was obtained from Malayalam (target language). The subjects who

underwent training were from three different language groups, bilingual Bengali-English and

Spanish-English speakers and monolingual American English speakers. With the description of

each language, its distinctive segments are discussed with their phonetic allophones. For the

proposed experiment, description of phonemic inventories alone would be insufficient to capture









the initial perceptual exposure as well as the developmental acquisition of the novel non-native

speech contrasts. The comparison of all four languages in question regarding the retroflex feature

is relevant to the assessment of factors such as feature generalization and/or multilingual benefit

since exposure to phonemic or allophonic experience with the retroflex feature may be either

beneficial or detrimental to the perception of a novel speech contrast (Polka, 1991; Strange,

1995; Pruitt et al., 2006). The study looks at the acquisition of Malayalam speech contrasts by

Spanish-English bilinguals, Bengali-English bilinguals and English monolinguals.

Malayalam

A member of the Dravidian language family, Malayalam uses the retroflex feature in most

of the manners it employs. This language was chosen to be the target language for the

experiment since the retroflex place of articulation is highly productive in this language.

Generally, retroflex sounds are taken as a specific place of articulation whereby a voiced

retroflex stop in one language would have similar articulatory characteristics in another

language. Contrary to this opinion, Ladefoged and Bhaskararao (1983) revealed that retroflex

speech sounds found in three languages, one Indo-Aryan and two Dravidian, displayed a varying

degree of retroflexion. Hey claimed that the degree of retroflexion and which place of

articulation the tongue blade touches could be compared with vowel heights in the way that it is

possible to make a continuously changing set of sounds from an apical dental to retroflex place

of articulation. In their study, Tamil, a Dravidian language closely related to Malayalam, showed

a high degree of retroflexion whereas Hindi, an Indo-Aryan language related to Bengali, showed

least degree of retroflexion. Malayalam, like Tamil, is assumed to have a high degree of

retroflexion.

For the purpose of the experiment, four speech contrasts were chosen from this language to

serve as training stimuli. The common attribute in all of the four contrasts is the presence of the









retroflex place feature in one of the segments in the contrast. The segments containing the

retroflex feature are novel speech sounds for the listener groups. Table 2.1 shows the phonemic

and phonetic segments of Malayalam (adapted from Asher, R.E. and Kumari T.C. (1997)).

Table 2-1. Phonemic consonant inventory of Malayalam (adapted from Asher, R.E. and Kumari
T.C. (1997).
Labio-
Bilabial Dental Alveolar Retroflex Palatal Velar Glottal
dental
Nasal m n n ra rR
Approximant v 41 j
Trill/ Tap r, r
Fricative s S h
Lateral approximant 1 1

Nasals

Malayalam exhibits a rich inventory of nasal stops employing a six-way place contrast

starting with bilabial, dental, alveolar, retroflex, palatal up till the velar. Researchers agree on

most of the phonemic distinctions among the different place contrast except for the dental -

alveolar place contrast. The dental alveolar place contrast has been discussed at length

describing its phonological as well as phonetic distinction in the language (Kumari, 1972;

Sreedhar, 1972; Kalackel, 1985; Raja, 1960; Mohanan and Mohanan, 1984; Asher and Kumari,

1997).

On articulatory grounds, this particular place contrast is marked universally in the world's

languages, since dental and alveolar place contrasts have very subtle distinctions articulatorily

and acoustically. It has been described in various ways on the basis of its articulation and

provided with phonemic pairs; terms such as 'spread-contact vs. point contact (Aiyar, 1972),



1 There is a disagreement among researchers about the phonetic properties of this sound. Asher and Kumari (1997)
refers to it as a sublamino-palatal approximant (discussed further in this chapter).









apico-laminal dental vs. apico-alveolar (Asher and Kumari, 1997) or simply dental vs. alveolar

(Subramoniam, 1973) have been used for this place contrast.

On phonological grounds, this place contrast's productivity in Malayalam is limited to only

one kind of distribution as geminates occurring medially. Due to this functionality, dental nasal

and alveolar nasal are considered phonemic in Malayalam (Kumari, 1972; Kalackel, 1985; Asher

and Kumari, 1997). However, Mohanan & Mohanan (1984) noted that dental and alveolar nasals

were perceived as separate in word-initial as well as intervocalic positions in nonsense words by

native Malayalam speakers, which perhaps implies that native speakers of Malayalam can

distinctly perceive the dental and alveolar nasals even in distributions not found in Malayalam.

On the other hand, apart from the geminate-occurring-medial position, the rest of the

distributions in Malayalam for this contrast appear to be complementary to each other. The

distribution for this contrast is described as [n] occurring non-finally and [n] occurring non-


initially (Kumari, 1972; Sreedhar, 1972). In addition, the orthographic script of the language

does not distinguish between the dental and alveolar place contrast (Asher and Kumari, 1997).

For the purpose of this experiment, dental and alveolar nasals in Malayalam are assumed to

maintain phonemic distinctions. Of particular interest for this experiment is the distribution of

this place contrast in intervocalic position. Unlike the dental nasal, the alveolar nasal occurs in

intervocalic position (Raja, 1960; Kumari, 1972; Sreedhar, 1972; Kalackel, 1985; Asher and

Kumari, 1997). The training stimuli employ VCV as the only syllabic position. Therefore, the

speech contrast [n] [rE] results in being produced as alveolar nasal vs. retroflex nasal since


intervocalically the nasal is realized as alveolar. This assumption does not conflict with Mohanan

& Mohanan's (1984) observation that Malayalam speakers can hear and reproduce the dental and

alveolar nasals distinctly in an intervocalic position. The perception of these segments in new









phonetic contexts cannot be interpreted to mean that speakers can produce dental nasal

intervocalically without any auditory prompt. Also, the reproduction of segments correctly in

Mohanan & Mohanan's observations may have been due to short-term memory effect.

Therefore, for the current experiment the contrast alveolar retroflex nasal was used as one of

the training stimuli.

Fricatives

The voiceless fricatives in Malayalam do not have voiced counterparts. The fricatives in

question, namely the lamino palato-alveolar [S] and retroflex fricative [f] were introduced in

Malayalam through Sanskrit loan words (Asher and Kumari, 1997). However, these fricatives in

Modem Malayalam have been so well integrated in the language that they're considered native

to the language (Kalackel, 1985; Sreedhar, 1972 etc). The place contrast of palato-alveolar -

retroflex fricative ([S] [f]) was employed as one of the training stimuli from Malayalam.

Laterals

Laterals were chosen as one of the stimuli speech contrasts due to the fact that the lateral

approximants in Malayalam exhibit a two-way place contrast, alveolar and retroflex. This place

contrast most appropriately measures up to nasal or fricative place contrasts in terms of

homogeneity of place of articulation as compared to the rhotic place-manner contrast in the

present experiment's stimuli. Moreover, the retroflex lateral approximant is uniquely

characteristic of Dravidian languages. This provides an opportunity to investigate the acquisition

of this place contrast by Indo-Aryan languages including Bengali which do not maintain

alveolar-retroflex lateral approximant place contrast.









Rhotics

Most researchers talk of another lateral/ continuant sound which is considered unique to

the Dravidian language family orthographicc symbol: YP). However, there is disagreement to a

large extent on its phonetic properties. It has been described in Malayalam as

Voiced retroflexed palatal fricitivised lateral [1] (Kumari, 1972)
Retroflex lateral fricative (no contact at the post-alveolar region) [1] (Sreedhar, 1972)
Voiced sublamino-palatal approximant [z] (Asher & Kumari, 1997)
Voiced apico-prepalatal approximant [1] (Kalackel, 1985)
Velar retroflex frictionless continuant [1] (Raja, 1960)
Retroflex approximant [1] or [z.] (Bright, 1998; Krishnamurti, 2003)

In articulatory terms, Kalackel (1985) describes it as

The apex of the tongue is grooved as if for pronouncing /2/, but it is bent back so that the
tip of the tongue is pointed at the root of the mouth where it is highest. The voiced sound
which is thus produced is not actually a lateral (since the air escapes not over the sides of
the tongue, but over the tip).

Clearly, a detailed phonetic description is required for this peculiar segment. For

convenience, the term retroflex approximant along with the IPA symbol [ t ] is used as reference

to this particular speech sound in this study. This speech sound is contrasted with alveolar tap [r]

to create a paired contrast with rhotic properties.

The case of alveolar tap retroflex approximant ([r] [4]) speech contrast is more complex

than other alveolar/palato-alveolar retroflex speech contrasts, in that the speech sounds contrast

in manner of articulation as well as place of articulation. This speech contrast is not homogenous

with other speech contrasts chosen in terms of contrasting attributes between speech sounds.

In all, four speech contrasts were chosen from Malayalam to examine the acquisition of

novel non-native speech contrasts by three listener groups. The novel speech sound in all of the

contrasts is the one containing the retroflex place feature.









The following subsections deal with the three listener languages, Bengali, Spanish and

American English. These languages will be discussed focusing on the presence or absence of

retroflex features in their phonemic or phonetic segments.

Bengali

Bengali, an Indo-Aryan language, displays productiveness of the retroflex feature in

manners like plosives and flap. It was chosen as one of the listener languages. Unlike

Malayalam, Bengali lacks the alveolar retroflex place contrast in manners like nasals, laterals

and fricatives (Bhattacharya, 1988; Ferguson and Choudhary, 1960; Kostic and Das, 1972;

Tunga, 19952; Ray et. al., 1966). These gaps in the consonant inventory make Bengali an

appropriate choice of language to investigate the acquisition of alveolar-retroflex place contrasts

in these manners. In addition, comparing this language group (Bengali-English bilinguals) with

the rest of the listener language groups allows us to investigate feature generalization of the

retroflex feature.

Bengali has a three-way place contrast in nasal stops, namely bilabial, alveolar and velar.

Unlike north-western Indo-Aryan languages (E.g Hindi), Bengali has lost the retroflex place

feature in the nasal manner of articulation, although the Bengali script retains the orthographic

symbol for the retroflex nasal. Therefore, one phoneme /n/ is used for two graphemes (q,'T)

(Bhattacharya, 1988). However, Bengali speakers have phonetic experience of the retroflex nasal

as an allophone. When /n/ precedes a voiced retroflex stop in a consonant cluster, it undergoes

assimilation and is realized as a retroflex nasal (Ferguson & Choudhary, 1960; Kostic & Das,

1972; Tunga, 1995). Although Polka (1991) and Strange (1995) noted that phonetic experience



2 Tunga (1995) noted that Midnapore dialect of Bengali (bordering the Oriya language area) uses the retroflex nasal
[ri] more often than the dental nasal [n]. None of the subjects spoke this dialect.









of a sound can influence the perception of a non-native speech contrast, it remains to be seen if

the retroflex feature with varying degrees of retroflexion can have similar effect due to exposure

in allophonic experience.

The degree of retroflexion in Indo-Aryan languages has been observed to be much less

than that of Dravidian languages (Ladefoged and Bhaskararao, 1983). The retroflexion in

Bengali is claimed to have a slighter degree of retroflexion than Malayalam (Chatterji, 1921;

Kostic and Das, 1972; Haldar, 1986; Tunga, 1995).

Table 2-2. Phonemic consonant inventory of Bengali
Labial Post-dental Alveolar Retroflex Alveo-palatal Velar Glottal
p ph t th t tS tSh k kh
Stops b bh ddh h d3 dh h


Fricatives s S h
Nasals m n (r) r)
Tap/flap r f
Lateral
approximant
* Adapted from Bhattacharya, K. (1988)

The fricatives in Bengali have similar story. Table 2-2 displays the phonemic inventory of

Bengali. Phonemically, Bengali displays palato-alveolar [S] and dental [s] fricative. The palatal


fricative occurs in all syllabic positions as well as intervocalically as geminates. The dental

fricative however, can only be realized in consonant clusters positioned word-initially or occurs

in loan words (Bhattacharya, 1988; Ferguson and Choudhary, 1960; Kostic & Das, 1972; Ray et.

al., 1966; Haldar, 1986). Thus, the palatal place contrast is the most productive in fricative

manner of articulation. Bengali has lost the retroflex [f] fricative originally derived from

Sanskrit. However, as with the nasals, the Bengali script retains the orthographic symbol for the

retroflex fricative. There are three graphemes (4T,4`,5) for a single phoneme /S/.









Bengali maintains a single place in the case of lateral approximants, alveolar /1/

orthographicc symbol: ff). It lacks the retroflex place of contrast in laterals as a distinctive

segment. However, Bengali speakers do experience retroflex lateral [1] as an allophone of/l/ in


certain environments. Alveolar laterals preceding retroflex consonants undergo assimilation and

are realized as retroflex lateral approximants, especially in consonant sequences like /1C/ or /lc/


(Bhattacharya, 1988; Ferguson and Choudhary, 19603; Tunga, 1995).

Rhotics in Bengali have a two-way place contrast, namely alveolar I/r and retroflex /1/


flap. The alveolar flap occurs in medial and word-final positions. Alveolar trill [r] is realized as

an allophone of this phoneme when it occurs word-initially. Ferguson & Choudhary (1960)

observed that sometimes an alveolar flap occurring after a labial or retroflex stop gets realized as

an approximant (like English [j]). Similarly, the retroflex flap when occurring in final position or

followed by a consonant sometimes is realized as a prolonged retroflex continuant. Therefore,

Bengali speakers may have allophonic experience of retroflex approximant as well.

Nevertheless, none of the Malayalam retroflex speech sounds being tested for acquisition occur

phonemically in Bengali.

Spanish

Spanish, one of the Romance languages, uses the palatal place of articulation extensively.

Most Spanish dialects allow manners of articulation like nasal, fricatives, approximants and

laterals at the palatal place of articulation (Hammond, 2001). This language was chosen as

another listener language involving the Spanish-English bilingual group, based on the


3 Ferguson & Choudhary (1960) also claims that a sequence like/ 1/ (where alveolar lateral follows a retroflex
consonant, both consonants result in a long retroflex lateral [1,].









nonexistence of the retroflex feature in its phonemic as well as phonetic segments. The lack of

experience with retroflexes makes the Spanish speaking subjects ideal to test the effects of meta-

linguistic awareness since any kind of linguistic experience with retroflex is controlled in this

language group. Table 2.3 shows phonemic consonant inventory of Spanish for the relevant

manners of articulation, namely rhotics, nasals, approximants and lateral approximants.

Table 2-3. Phonemic consonant inventory of Spanish (Only manners that are relevant for the
study are displayed here.)
Bilabial Labio-dental Dental Alveolar Post-alveolar Palatal Velar
Trill r
Tap r
Nasal m n pi
Fricative [13] f 0 [9] s x [ y]
Approximant j w (labio)
Lateral 1 K
approximant
*Adopted from: Hammond (2001)

Rhotics in Spanish are considered another unique attribute of Romance languages. The

alveolar trill is phonemically contrasted with the alveolar tap (Lindau, 1985). The alveolar trill is

not found as a phonemic sound in either English or Bengali. In the case of the alveolar tap, while

being a phonemic consonant in Bengali, it occurs only in phonetic relationship with alveolar /t/

and /d/ stops in American English (Roach, 1983; Olive et. al., 1993; Ladefoged and Maddieson

1996).

Nasals as well as lateral approximants in Spanish have a two-way place contrast described

as alveolar and palatal. This contrast is prevalent in most of the South American dialects of

Spanish. In the case of fricatives, most of the Spanish dialects maintain the alveolar place

contrast in sibilants and do not have an equivalent of palatal or palato-alveolar fricative; the

exception being Argentinean Spanish where in the fricative manner a palatal allophone also

occurs along with the alveolar one.









There appears to be complete lack of retroflex features in the sound inventory of Spanish

whether phonemic or phonetic. Therefore, the acquisition of retroflex sounds or specifically the

alveolar-retroflex place contrasts from Malayalam may pose a degree of difficulty, attributed to

acquisition of new non-native speech sounds, for the speakers of Spanish.

American English

American English was included as the third listener language for the experiment. This

listener group (monolingual American English group) acted as the control group for the

experiment. English employs the alveolar place of articulation extensively to create phonemes in

different manners of articulation. Table 2.4 shows the phonemic consonant inventory of

American English of the relevant manners of articulation.

Table 2-4. Phonemic consonant inventory of American English (Only manners that are relevant
for the study are displayed here.)
Bilabial Labio- Dental Alveolar Post- Palatal Velar glottal
dental alveolar
Nasal m n rQ
Fricative f v 0 0 s z S 3 h
Approximant a j w(labio)
Lateral 1
approximant
*Adapted from: Ladefoged (1999)

The lateral approximant in American English employs only the alveolar place. The

phoneme /1/ has a velarized lateral approximant allophone, syllable-finally. Nasals in American

English have a three-way place contrast namely, bilabial, alveolar and velar. The post-alveolar

fricative, similar to palato-alveolar fricative in Malayalam, does exist in American English.

However, the langauge lacks any existence of retroflex fricative.

The alveolar tap, while being a phonemic consonant in Malayalam, occurs only in a

phonetic relationship with alveolar /t/ and /d/ stops in American English (Ladefoged and

Maddieson, 1996). In the case of retroflex feature exposure, unlike Spanish, English speakers









may have some phonetic experience with the retroflex feature in incidental contexts. Polka

(1991) noted that English speakers may have some experience of retroflex stops at the phonetic

level. The alveolar stops /t/ and /d/ may be produced as retroflex stops /t h/ and /ct /when the


alveolar stops precede /J/ in words like "try" and "dry". However, as noted by Polka (1991), this


is not a predictable phonological process and may vary on a dialectal or individual level. Also, it

should be noted that the consonants in question (/1, n, r, sh/) do not precede the rhotic

approximant in syllable-initial position in American English. Therefore, AE speakers do not have

the experience of their becoming retroflexed as is the case with /t, d/.

Stimulus Materials

The novel non-native speech contrasts for the study were taken from Malayalam. It was

chosen as the target language as it has a rich inventory of consonants employing the retroflex

feature extensively in various manners. This allows a flexible choice of speech contrasts as per

the requirements of the study (as discussed below).

The speech contrasts chosen for the study are:

(1) [ 1 ] (alveolar versus retroflex lateral approximant)
(2) [n r. ] (dental/alveolar versus retroflex nasal)
(3) [J- ] (palato-alveolar versus retroflex fricative)
(4) [ r -4 ] (alveolar tap versus retroflex approximant)

These speech contrasts differ primarily in place of articulation, although the [ r 4 ]

contrast also differs in manner. The reason for choosing these particular speech contrasts, when

looking at the Bengali-English (BE) language group, is primarily for the purpose of evaluating

feature generalization. One of the speech sounds in each speech contrast from Malayalam is

marked with the retroflex feature [ [, r, 4 ]. These sounds are not found in Bengali which

employs the retroflex feature but only in the oral stop and flap manners of articulation.









Therefore, if the retroflex feature gets generalized from existing phonetic categories to newly

formed phonetic categories, the evidence of any variability in generalization of the feature in

different manners will be reflected during the acquisition of these four new speech contrasts.

In case of the monolingual language group, American English (AE) speakers do not have

experience with retroflex sounds phonemically. However, retroflexion of alveolar stops in

particular environmental contexts (e.g. [ 4j&o] 'dry' (Polka 1991)) is found in the speech of

American English speakers. Spanish-English (SE) bilinguals have no experience with the

retroflex feature in phonemic or phonetic environments when using Spanish. However, being

bilingual, they are assumed to have phonetic experience similar to American English speakers.

Having no phonemic exposure to retroflexion makes the Spanish-English bilingual group ideal to

examine the multilingual benefit.

The materials were recorded as natural stimuli by Malayalam talkers producing each of the

eight speech sounds from the four speech contrasts discussed above. Each speech sound was

embedded between two identical vowels (vowel contexts: [a, i, u]) resulting in a set of non-

words. Six repetitions of each consonant with each vowel context were recorded in order to use

the best possible tokens for the stimuli.

Recording for the stimuli was done in India (Chennai, Chandigarh, Bangalore, Cochin, and

Ahmadabad). Currently, all speakers that were recorded have been residing out of the state of

Kerala. Ten native speakers of Malayalam were recorded (See figure 2-1 for their region

distribution). However, due to errors in elicitation, two speakers (M03 and M05) were not

considered for this study. Detailed information of the Malayalam speakers is provided in table

2-5.









Table 2-5. Talker information; all talkers are native speakers of Malayalam and born in the state
of Kerala, India; Years= "years not lived in their home state".
Talker Code Gender Age Years Other Languages
known
M01 F 28 4 Kannada
English

M02 F 37 1 Hindi
English

M04 M 45 15 Hindi
Punjabi
English

M06 F 36 9 Hindi

M07 M 28 1 English
Tamil
M08 M 29 5 Hindi
English

M09 M 32 8 Hindi
English

M10 M 29 5 Hindi
English

The data were recorded and digitized at 44.1 KHz on a Marantz Digital recorder and

transferred directly onto a personal computer. The recordings took place in a quiet room with no

significant reverberation or background noise. The subjects were provided with the non-words

(in Malayalam script) on the computer with an inter trial interval (ITI) of 2.5 seconds. The

talkers were asked to read through the list before the recording in order to be familiar with the

non-word materials. The data were segmented using PRAAT speech analysis software.




































Figure 2-1. A map of the state of Kerala with cities of origin for the Malayalam speakers. A
smaller map of India on the left hand bottom corner shows the location of the state of
Kerala.

Stimulus Material Evaluation

Before developing the stimulus material for the main study, the recorded data from the

Malayalam talkers was evaluated. Native Malayalam listener ratings were elicited for tokens

from eight talkers. Four native speakers of Malayalam evaluated the intelligibility of the

recorded stimuli through a consonant identification task. Tokens that were identified consistently

as the target sound 83% 100% were selected for the stimuli. Three out of the selected tokens,

closest in duration and mean fundamental frequency, were selected for each consonant with all

three vowel contexts, making a total of 72 stimuli (8 consonants x 3 vowel contexts x 3 tokens)

per talker.









Stimuli Evaluation: Preliminary Experiment

Conducting the main experiment would have been unproductive for the purpose of

evaluating the role of assimilation if ceiling or floor effects were observed for any of the speech

contrasts in the monolingual language group (AE). Therefore, to avoid this possibility, a

preliminary experiment was conducted to examine the discrimination level of these contrasts by

monolingual American English speakers since the AE monolinguals acted as the control group in

the main experiment, against which the results of the bilingual groups are to be compared. Thus,

it is considered reasonable to test the difficulty level of the stimuli with the control group.

The goal of the preliminary experiment was to examine the level of discriminability of the

novel non-native speech contrasts by American English speakers. Eleven monolingual American

English subjects were recruited for this experiment. All were undergraduates at the University of

Florida. They were compensated with extra credit or paid cash ($5/40 minutes). None of them

reported average proficiency in any language other than American English.

For stimuli, two tokens of each non-word (8 consonants x 3 vowel contexts) per talker

were used. Data from three talkers was used. In all, there were 1152 stimuli pairing up to make

576 trials (16 trials per speech contrast different trials, identical trials and physically different

trials).

Subjects were given a categorical AX discrimination task. (Please refer to the section later

in this chapter for a description of an AX discrimination task.) Because physically different

tokens of the stimuli are used, the discrimination task has to be administered as a categorical

task. Naturally produced stimuli may have subtle irrelevant acoustic differences which may

influence subjects' decision away from category-level processing. Therefore, the subjects are

asked to decide whether both stimuli sound like they are from same category or different









category. For this task, the inter stimulus interval (ISI) was provided as 1 sec and inter trial

interval (ITI) was 4 sec.

The results of the preliminary experiment were as follows. Each subject's performance in

the AX test was converted to a d'score(scale of 0 to 4). Table 2.6 shows the results of the

discrimination task for the American English speakers.

Table 2-6. Result of the preliminary AX discrimination task across three Malayalam speakers.
Mean D prime values averaged across contrasts as well as across vowels.
[1-1] [n-] [r-If [S-] Average across contrasts

[a] 1.09 2.55 2.99 2.70 2.33

[i] 0.94 1.11 1.69 2.28 1.51

[u] 1.37 1.25 2.89 1.09 1.65
Average across vowels 1.13 1.64 2.53 2.02


The [r-4] contrast averaging over all three vowel context was the most discriminable of the

four speech contrasts (d" value 2.5). Overall, the [ 1-[ ] speech contrast was least discriminable

(1.1) followed by the nasal contrast [ n-r ] (1.6) and fricative contrast [ S- ] with devalue at

(2.02). Among the vowel contexts, the [i] context was least discriminable (1.5) closely followed

by [u] (1.6) whereas [a] was the most discriminable (2.3). Across the four speech contrasts, the

[i] vowel context was the least discriminable (with the exception of fricative [ JS- ] contrast

where [u] was the least disciminable) and [a] was the most discriminable. No ceiling or floor

effects were found for any of the non-native speech contrasts.

Stimulus Preparation

Based on the results of the preliminary experiment as well as native listener ratings, the

stimulus materials for the perception training experiment were selected. For each talker, three

physically different tokens per consonant and vowel context (8 consonants x 3 vowel contexts x









3 tokens) were chosen. The tokens of one talker (M08) were used solely for the familiarization

task. Talker M04 was used for pretest and posttest as well as discrimination test and perceptual

assimilation test, since all of these tests have to be identical in stimuli and procedure in order to

be directly compared. M04 was also included during training since the testing pretestt and

posttest) is on a trained talker. M07 was used for the generalization test only. This would be the

novel talker, whose voice was not exposed to the subjects anytime during the pretest or training

phase. Data of six talkers (M01, M02, M04, M06, M09, M10) was used for training sessions.

Familiarization phase: one talker (8 consonants x 3 vowels x 2 tokens x 2 repetitions) 96
trials in all.

Pretest/posttest phase

o Pre-test/post-test: one talker (8 consonants x 3 vowels x 2 tokens x 6 repetitions)
288 trials in all.

o Discrimination test: one talker (8 consonants x 3 vowels x 2 tokens x 4 orders)
192 trials in all.

o Perceptual assimilation test: one talker (8 consonants x 1 vowel x 2 tokens x 5
repetitions) 80 trials in all.

o Production pretest/posttest: one talker (8 consonants x 3 vowels x 1 token x 6
repetitions) 144 trials in all.

o Generalization test: one talker (8 consonants x 3 vowels x 2 tokens x 6
repetitions) 288 trials in all.

Training phase: two talkers per session (8 consonants x 3 vowels x 3 tokens x 2
repetitions) 288 trials per session. Over a total of six training sessions, data from six
talkers is used.

Sound-symbol association issue: The issue of sound-symbol association was foreseen in

the design of the training experiment. Since the subjects (all three language groups) are not

familiar with Malayalam orthography or vocabulary, it would be difficult to train them without

sound-symbol association. This kind of issue has not arisen in the previous studies (Logan et al.,

1991 learning of/r/-/1/ distinction by Japanese learners of English) as the subjects were familiar









with the vocabulary and alphabets of the target language. In the current study, the question of

orthography associated with the 8 consonants was resolved by using arbitrary symbols for each

consonant. The arbitrariness of the symbols being associated with the sounds is inevitable since

these are novel non-native sounds for the listeners.

A number of alternate methods of representing the sounds were considered before

appropriate symbols were finalized. Firstly, using the original Malayalam script or IPA symbols

with the sounds would have taken the subjects much more time and effort for the sound and

symbol association resulting in decreased focus on the perception of sounds. Secondly, purely

arbitrary geometric shapes as symbols were dismissed as potential candidates for the same

reason as above. In addition to that, the confounding possibility of learners perceiving the sounds

correctly but clicking on the wrong geometric shape could be detrimental to the validity of the

results. The third alternative was using the English alphabets as representative symbols. The

option was considered as the subjects in all the language groups were familiar with English

alphabets and their corresponding English consonants. The small and upper case of the alphabet

'n' could be used for the [n] [qj sound distinction so that whenever subjects hear a nasal, there


would not be the basic confusion of finding the correct symbol out of the 8 symbols, as would be

the case with purely arbitrary representations. One caveat with this type of representation could

be that the subjects may not easily learn the distinction between the small and upper case letters

representing the alveolar-retroflex distinction. Therefore, to avoid this possibility, all

dental/alveolar sounds in the four sets of contrasts that were presumed to be assimilated to their

L1/L2 category were assigned the small case letters. The retroflex sounds in all the four sets of

contrasts were assigned the corresponding upper case letters. In addition to this, during

familiarization process, the sounds in each set were auditorily prompted (along with their









corresponding symbol) next to each other. The subjects were told that the sounds they will hear

are not English sounds but from a language not known to them.

After careful consideration, the orthographic representation of the sounds was finalized as

small and upper case English alphabets. Hence, the study uses the following symbols in training:

1= dental lateral approximant; L=retroflex lateral approximant; n=dental nasal; N=retroflex nasal;

sh=palato-alveolar fricative; S=retroflex fricative; r=alveolar tap and R=retroflex approximant.

These arbitrary symbols were familiar to all three language groups and their efficacy was

verified in pilot testing.

Subjects

The study employed a diverse range of challenging non-native consonant contrasts from

Malayalam with three populations of listeners,

* Bengali-English (BE) bilingual group
* Spanish-English (SE) bilingual group
* American English (AE) monolingual group

Twenty (20) participants were recruited for each group from the University of Florida

campus. At the time of the experiment, the participants, ages ranging from 18 to 35, were

enrolled in undergraduate or graduate courses at the university.

For all subjects, the level of proficiency was self-reported through a language background

questionnaire. Self-reported language proficiency may not be an accurate measure for screening

of potential subjects for the study. Therefore, certain precautionary measures were taken while

recruiting the subjects for the three language groups. In the case of American English

monolinguals, the experimenter asked for specific information regarding the knowledge of other

languages or extent of exposure to other languages apart from English (LI). The preliminary

(introductory) email sent out to the interested individuals has been provided in Appendix A. If









the candidate had taken Spanish language classes in middle/ high school, which was mostly the

case, a brief test of spoken Spanish was conducted to assess the level of LI transfer from

English. The procedure and content of the spoken Spanish test has been explained in Appendix

A. In the case of Bilingual Spanish-English subject recruitment, in addition to conducting the

language background questionnaire, the experimenter assessed the level of proficiency in spoken

American English. These individuals were also asked to take the spoken Spanish test. The

screening for bilingual Bengali-English subjects involved assessing their level of spoken

proficiency in English by observing any noticeable signs of LI transfer in consonantal segments

during conversation. More specifically, the Bengali-English participants were screened for any

exposure to Dravidian languages and if present then to what extent. Candidates with even

minimal awareness of retroflex sounds were not recruited for this study.

During the screening process, bilinguals, preferably early bilinguals with an age of

acquisition (AOA) for L2 of fewer than 12 yrs, were considered as potential subjects. If age of

acquisition (AOA) for L2 was more than 12 yrs, then candidates with self-reported proficiency

level of L2 as HIGH were considered as potential subjects. Specifically for Bengali-English

subjects, the English refers to Indian English since most of the student population is graduate

students staying in U.S no more than 5 years. Subjects' proficiency in English was self-reported

and also assessed by the experimenter.

For the monolingual subjects, self reporting of any language exposure was required

including conversational ability in another language or passive exposure to another language.

Subjects with very minimal knowledge of another language (mostly Spanish from High School

language classes< 2 years) were considered in the monolingual American English language

group. It was almost impossible to find pure monolingual subjects with no exposure to another









language especially in the age group of 18-35 years due to the educational curriculum in North

American Schools. None of the participants had any prior exposure to Malayalam.

In addition to this, subjects were asked in the screening questionnaire (Appendix A) of any

hearing or speech problem that they may have had in the past. All subjects reported normal

hearing with no hearing or speech impairment history. No preliminary clinical screening for

normal hearing was conducted. Therefore, the information collected on normal hearing was self-

reported.

Bengali-English Bilinguals

The Bengali-English language group was chosen for their phonemic experience with the

relevant place feature retroflexx in oral stop and flap manners of articulation), though Bengali

and English lack any direct correspondents to the Malayalam contrasts under study. The results

from the BE language group when compared with Spanish-English bilinguals were expected to

provide evidence for the feature generalization factor in the learning of a third language.

Twenty (20) speakers of Bengali with English as L2 participated in the study. The mean

age of participants in this group was 28 years. All subjects were from Kolkata, the capital of

West Bengal state with the exception of two subjects. BE16 and BE17 are from Barddhaman, a

southern district in West Bengal, close to Kolkata. They speak the standard colloquial Bengali,

similar to what is spoken in Kolkata. Most of the subjects have acquired a working knowledge of

Hindi through entertainment media, but only four subjects, BE03, BE05, BE1 land BE12, have

had formal education in Hindi. Some subjects have been exposed to languages like Assamese,

Oriya, Urdu, Nepali and Telugu through the surrounding environmental context from a few

months up to two years. However, even low proficiency in comprehension or speaking in these

languages was not reported. Therefore, any influence from the above mentioned languages can

be ruled out. The details of all the subjects have been provided in the appendix A.









Spanish-English Bilinguals

The Spanish-English (SE) bilingual group was chosen for their lack of linguistic

experience with retroflex place feature. The results from SE language group when compared

with other groups were predicted to provide evidence for effects of multilingual benefit in

bilinguals.

Twenty (20) speakers of Spanish with English as their L2 or another LI, participated in

this study. The mean age of subjects within this language group was 20 years. All subjects have

lived in North America all or most of their lives. In cases where subjects were born outside of

North America, they had shifted to North America within the first 6 years of their lives.

Therefore, these bilinguals acquired Spanish at home as their first language and picked up

American English either simultaneously with Spanish or during preschool years (age 3 to 6). The

dialects of Spanish spoken by these subjects are of Central American origin with the exception of

one subject who spoke Argentinean Spanish. Figure 2-2 shows the origins of the Spanish dialects

of the SE bilingual subjects. The table providing detailed information on individual subjects is

given in appendix A.













SE04
SE09
SEll
SE18


SE08
SE15
SE16
SE19


SE13
SE21


El Salvador -
SE03
SE12
SE02 .Pan
SEO0
SEI SE06
SE07
SE14
SE20


Figure 2-2. An outline map of Mexico, Central America and South America showing the
regional distribution of the subjects.



American English Monolinguals

American English speakers were recruited for the monolingual language group acting as a


control group for the experiment. The subjects (20.2 years as the average age within this


language group) acquired American English as their first language. None of the subjects reported


more than minimal proficiency in any language other than American English. Since most of the









students had lived most of their lives in Florida, many of them had to take Spanish as a required

language class. This factor could not be controlled for therefore, subjects were asked to read out

two brief tongue twister phrases in Spanish. This was done to determine if the subject had

actually acquired any perceptual categories corresponding to Spanish language. Only those

participants were recruited as subjects whose speech did not provide any evidence of Spanish

categories but in fact displayed transfer from American English. The Spanish phrases as well as

detailed information on individual subjects are provided in Appendix A.

Experimental Design

Perception Training Method

The present experiment used a perception training method with identification paradigm

containing high variability stimulus (High-variability perception training). This training emerged

as an effective training method in learning non-native speech contrasts through the academic

disagreement between two schools of thought regarding the mental representation of phonetic

sounds in the minds of L2 learners: the prototype model (Strange and Dittmann, 1984; Jamieson

and Morosan, 1986, 1989; Kuhl, 1991, 1992) and the exemplar-based model (Jamieson and

Morosan, 1986; Logan et al., 1991; Lively et al., 1993, 1994; Bradlow et al., 1997). In order to

comprehend the reason for adopting High-variability perception training method for this study, it

is imperative to discuss the theoretical assumptions underlying these two models of long term

memory representation.

The earlier school of thought, the prototype model, claimed that the phonetic sounds

acquired by a listener were represented as context-invariant units of sounds, the content of which

was the ideal example of that phonetic sound, otherwise named as the prototype, or in a

phonologically-driven view, phoneme. Every representation of a phonetic category in the mind

of an L2 learner was assumed to have foci, the most representative member of the category









(Kuhl, 1991, 1992). The foci was expected to have more stable representations in long-term

memory, since only linguistically relevant information was stored in that prototype and could

overlap concrete natural members of that category when perceived. However, empirical testing

of this training type failed to prove effective in developing long term memory representation

since the subjects in these studies failed to generalize to new phonetic environments. A limitation

in the notion of prototypical representation was that it was assumed that only linguistically

relevant information will be extracted to form the representation of the phonetic category.

However, by postulating this, the model assumed that other detailed information in those stimuli,

such as talker's voice, speaking rate or the phonetic environment, is filtered out. This detailed

information was established to be of relevance in the representation of phonetic categories by

subsequent empirical results (Logan et al., 1991; Lively et al., 1993).

The results from Logan et al. (1991, 1993), Lively et al. (1994) and Bradlow et al. (1997)

clearly indicated that for generalization to occur, detailed information like talker variability and

phonetic environments are as relevant as other linguistically significant information. The

perceptual learning and development of stable representations of phonetic categories result from

extensive stimulus variability in natural stimuli. This led to the formulation of the exemplar

model of phonetic representation, extending the exemplar-based model of categorization

(Nosofsky, 1986, 1987; Kruschke, 1992) based on the selective attention weighing mechanism.

Multidimentional representations are stored in the memory during training while selective

attention weighs the importance of various stimulus dimensions. Changes in selective attention

'stretch' or 'shrink' the perceptual space for these dimensions and alter the internal category

structure resulting in phonetic categories being less similar to other members as dimensions are









stretched and more similar to other members when dimensions are shrunk (as cited in Lively et

al., 1993).

The choice of training method for the current research is clear from the above discussion.

The present experiment used a perception training method with identification paradigm

containing high variability stimulus (High-variability phonetic training). The choice of

perception over production or pronunciation training is also based on the debate of mental

representation of phonetic categories in long-term memory. Since production/pronunciation

training is based on providing the learners with the ideal or best exemplar of the phonetic

category, the input lacks any variability at all (Dalby and Kewley-Port, 1999; Akahane-Yamada

et al., 1996; Neri et al., 2002 etc).

Procedure

A pretest-training-posttest design was employed, adapting the procedure used by Lively et

al. (1993, 1994). The effect of training was measured by comparing the performance in pretest

and posttest tasks. A consonant identification procedure was used throughout the training phase

of the experiment. Apart from the consonant identification task, an AX discrimination task and

perceptual assimilation task were also employed during the pretest and posttest phase of the

experiment. A minimum of seven-hour gap was used between sessions to avoid any confounding

effects of working memory influencing the perceptual performance before and after training and

also to prevent any adverse effects of fatigue during consecutive training sessions. The term

working memory, in this case, refers to immediate retrieval of information stored in the short

term memory, as a result of preceding training. On the contrary the results would be more valid

if the subject's response is based on their long term memory representation of sounds. This

arbitrary period of seven hours as the minimum gap between two consecutive train sessions was

considered assuming that the routine task of the individuals would negate any working memory









effects. All training and testing took place in a sound-attenuated room equipped with individual

computer stations (contains a keyboard, a CRT monitor and headphones) for subjects. Stimulus

materials were presented to subjects over a set of calibrated head phones. The software collected

individual responses during all phases of the experiment. All subjects were tested and trained

individually.

Familiarization

Subjects were familiarized with the symbols associated with each consonant. The talker

used for this task hyper-articulated the sounds during recording and, thus, was considered ideal

for the familiarization task. The stimuli consisted of one talker (8 consonants x 3 vowel contexts

x 2 tokens x 2 repetitions). In all, there were 96 trials and the ITI was 2 seconds. The sound file

was played while the symbol was displayed on the computer screen. The subjects were asked to

pay attention to both the speech stimulus and its corresponding symbol. They were told that the

words are from a language not known to them and the arbitrary symbols represent different

sounds. The non-words contrasting in place (eg. [ili] vs [iLi] )were presented within same vowel

context alternating between the two words for maximum impact of familiarization task.

Pretest

During the pretest phase, the perceptual performance of the non-native stimuli was

determined at the naive listener level through various tasks. Pre-testing of perception (consonant

identification task) was incorporated to test the level of accuracy in identification of the target

sounds. The stimuli for the pretest (identical task also given as posttest) consisted of one talker (8

consonants x 3 vowel contexts x 2 tokens x 6 repetitions), whose data is also used for the training

(Trained-on talker). In all there were 288 trials. The pretest consisted of a consonant

identification task presenting the randomized stimuli over the headphones. The sound file was

played and the response options became available for the subject to click on. The response









options provided were all of the eight symbols corresponding to the eight non-native sounds in

question, thus technically making 12.5% as chance level though no confusion between the

manners is expected since all LIs have dental/alveolar nasals, laterals and rhotics. The subject

had the option to change her/his response before going on to the next trial. Once the next trial

button was pressed, the answer was recorded by the software and no changes could be made

(same for all tasks). The task was self-paced and therefore no ITI was assigned. No correct

answer feedback was provided during pretest.

A categorical AX discrimination task was used to measure listener sensitivity to non-native

speech contrasts at the naive exposure level pretestt). The same task was administered during the

posttest phase so as to measure the perceptual performance on the same stimuli after training.

An AX discrimination task tests the subjects' level of perceptual performance on a given

stimulus set. Each trial consists of a set of two stimuli (either category-different or category-

same). Subjects are required to decide if the stimulus X (the second stimulus they hear) is the

same or different than stimulus A (the first of the stimuli set). The decision of similarity or

difference is required to be made on the basis of perceptual categories (phonemic sound, word

etc).

Two types of trials can be used in such discrimination task: physical identity trials and

categorical trials. The physical identity trials only pair up the same token for same trials so that

the trial represents the same exemplar of the speech sound. This provides insight into the

perception of purely physical properties of the sounds. On the other hand, categorical trials,

along with using physical identity trials, also employ the physically different tokens of the same

speech sounds as pairs for the same trials. This provides insight into the perceptual processing at

the phonemic level. For the discrimination task in this study, categorical trials were employed.









Twelve possible trials per speech contrast are usually employed. An example of these possible

trials is given in figure 2-3. In this test, the two same trials are repeated in order to maintain

equal number of same and different trials and therefore the number of same trials become four

making a whole set of 16 trials per speech contrast.

Inter stimulus interval (ISI), the time gap between the presentations of the stimuli in a

single trial, is also important in defining the difficulty level of the task. The longer the ISI, the

more difficult it is for the subject to rely the response only on working memory. Thus, phonemic

level processing comes into play and the longer ISI tests the perception at a categorical level.

Same trials Al A2 XI X2
A2 Al X2 XI
Different trials Al X1 XI Al
Al X2 X2 Al
A2 XI XI A2
A2 X2 X2 A2
Figure 2-3. Different trials possible in an AX discrimination task using two tokens of each
speech sounds.

Another task known as the perceptual assimilation task was administered during the pretest

phase to examine the similarity of the speech sounds presented as stimuli when compared with

the listener's native speech sounds. This test provides an insight into listener intuitions of

classifying the non-native speech sounds according to the already established phonetic categories

of the native languagess. In this study, the listeners were asked to listen to each stimulus sound

(provided with only one vowel context) and write down the perceptually closest speech sound (in

the native language orthography) in her native languagess. In case of bilingual groups, the

subjects were allowed to choose the response (the closest sound) from both LI and L2. They

were asked to write down the alphabet(s) from their native language corresponding to the sound

heard and specify the language (especially in the case of Spanish-English bilinguals since the









writing system of both English and Spanish overlaps substantially). This assimilation task was an

open-set test, where the listener was not bound by a few response options to choose from but was

allowed to choose from her past native linguistic experience.

Along with that, for each trial, listeners were asked to provide a category goodness rating

of how similar the native speech sound is to the newly heard stimulus sound. The goodness

rating scale consisted of 1 to 7 levels where 1 was to be chosen when the target sound was 'very

different' from the sound closest to it (according to the listener). Option 7 was to be chosen when

the target sound was 'exactly same' or in other words highly similar to the sound closest to it

(modal response for the token, according to the listener). Listeners were able to specify the

degree of difference between the sounds by circling a number between 1 and 7. The same task

was administered at the both the pretest as well as the posttest phase of the experiment in order to

evaluate any changes in assimilation patterns as an indication of developing new phonetic

categories. In all, the pretest phase consisted of a consonant identification test, an AX

discrimination test and a perceptual assimilation task.

Training

Subjects began the 6-session training phase after the pretest phase. The training procedure

used the same consonant identification task. The difference between the pretest task and the

training task is that feedback is given on every trial during the training sessions. If the response

was correct, the message 'you are correct!' was displayed. After 500 ms interval, the sound file

was replayed along with the correct symbol for reinforcement of the association. If the response

was incorrect, the message "Incorrect!!! Please listen to it again" was displayed. After 500 ms

interval the sound file was replayed and the correct symbol was displayed for reinforcement. The

subjects were not required to respond during repetition. The task was self-paced with no ITI

between trials. Data gathered from six talkers was used as stimuli for the training portion of the









experiment. Subjects heard only two talkers per training session. The sequence of presenting

talkers remained the same throughout training. Therefore, subjects heard each talker two times

during the 6-day training phase. The stimuli contained (8 consonants x 3 vowel contexts x 3

tokens x 2 repetitions) per talker. In all, there were 288 trials (2 talkers per session). The task was

paused for 60 seconds after one block of trials (stimuli of one talker). The subjects were allowed

to take breaks for longer time if required.

Posttest

A final posttest was given at the end of the training phase. The posttest consisted of the

consonant identification task, discrimination task and perceptual assimilation task (all three

identical to pretest). Another similar identification task was administered but with untrained

stimuli and one talker (novel talker) to measure the generalization (generalization test) of the

training to novel stimuli with new talker voice information. The design of the generalization test

was identical to the consonant identification test except that the stimuli were taken from a novel

untrained talker.

Task Order Presentation

The order of presentation for the various tasks during the pretest phase was kept same for

all the subjects across language groups. Once the familiarization task was over, the pretest phase

began with the consonant identification pre-training test as the initial task. It was followed by the

AX discrimination test. After a brief rest for five minutes, the subjects were given the perceptual

assimilation task. The presentation of the stimuli in each of the tests was identical for all the

subj ects.

The consonant identification test, AX discrimination test and the perceptual assimilation

task that were conducted at the pretest level were repeated at the posttest level of the experiment

as well. In addition to that, a generalization test was also conducted during the posttest phase.









The presentation order of these tasks was as follows: consonant identification post-training test

followed by the AX discrimination test and the perceptual assimilation task. Towards the end,

the generalization test, where perceptual performance on stimuli from novel talker was assessed,

was conducted.

Data Analysis

The results of the pretest and posttest phase were analyzed across contrast and vocalic

context as well as across language groups. The mean percentage scores for the consonant

identification tasks and d prime scores (discussed further in detail) for the discrimination tasks

were obtained. Results from the generalization test were analyzed across groups and individually

when compared with posttest results. ANOVAs were conducted separately for each contrast to

test the hypotheses under consideration.

The calculation of d prime scores and perceptual assimilation types are discussed herewith.

For the analysis of discrimination test results, the testing software captured '1' if the subject

clicked on the response item 'same' and '2' if the response item was 'different'. These responses

were then converted to a binary code, termed for this study as signal present. Hit rate, which

means that the subject correctly identifies the distinction within a contrast, was calculated by

averaging the signal present for all the 'different' trials. False alarm rate, which means that the

subject has marked all responses the same regardless of the trial type (either same or different),

brings out such discrepancies in the result and helps in accurately obtaining the d prime score.

The false alarm rate was calculated by averaging the signal present for all the 'same' trials.

Based on these, the d prime score was then calculated by subtracting the NORMINV value of

false alarm rate from that of the hit rate. NORMINV formula returns the inverse of the normal

cumulative distribution for the specified mean and standard deviation. Once the d prime values

were obtained for the pretest discrimination test as well as the posttest discrimination test, a









difference score (posttest value-pretest value) was obtained and later averaged vowel context per

contrast for each subject. These difference scores were submitted to the statistical tests for the

discrimination results.

The calculation of perceptual assimilation types from the data collected during the PA task

was as follows. The responses written by the subjects were recorded in excel files. To arrive at

the modal response for each sound, a matrix was developed which calculated the percentage of a

particular response for that speech sound. Once the modal response for each of the sounds was

reached at, the general assimilation types were obtained. The resultant general assimilation types

for each contrast (pre as well posttest values) were either within category (WC) or between

categories (BC). A contrast was BC if the modal responses for the sounds in that contrast were

different. On the other hand, a contrast was WC if the modal responses were same and the

overlap between them was less than or equal to 0.1 (maximum). For the between category

assimilation types, if the overlap was more than or equal to 0.9, it was termed as two category

assimilation (TC) since the overlap of modal responses was minimum. If the percentage count

for modal responses for both speech sounds within a contrast was less than 0.9, then it was

termed as both uncategorizable assimilation types (UU) otherwise it was termed as

uncategorizable-categorizable assimilation type (UC). In the case of within category assimilation

types, the ratings for the speech sounds in that contrast were submitted to TTest (statistical

measure) in order to examine any significance difference between the goodness ratings given for

the modal responses corresponding to each speech sound. If significance was found between the

ratings given, the contrast was termed as category goodness assimilation type (CG), otherwise

single category (SC). (Please refer back to Chapter 1 for detailed description of the various

assimilation types under Perceptual Assimilation Model.)









Summary

In this chapter, the methodology of the main experiment was provided. Starting with the

description of the target language (Malayalam) and three listener languages (Bengali, Spanish

and American English), the chapter provides details of stimuli preparation and evaluation as well

as issues related to stimuli presentation. Information on subject groups was given, including the

explanation to the appropriateness of the choice of the subject groups. A review of the perception

training paradigm involving discussion of theoretical issues in mental representation of

perceptual categories is considered essential for the understanding of the perception training

method chosen. Finally, a detailed procedural description of the experiment was provided.









CHAPTER 3
RESULTS

Introduction

For this study, three goals were put forth in chapter one: to investigate the factor of

multilingual benefit in cross-language speech perception; to examine the role of feature

generalization in facilitating the acquisition of non-native speech contrasts by trained /

experienced listeners; and to investigate changes in assimilation patterns of these speech

contrasts, if any, before and after training in individual language groups as well as across groups.

Of these, the third goal is descriptive and is discussed in the last section of this chapter.

The design of the experiment, mainly the choice of the subject groups, has its basis on the

hypothesis formed for this study. It is expected that a positive influence of multilingual benefit

on the acquisition of speech contrast will be seen in both bilingual groups through their better

performance in Identification and discrimination tasks when compared to monolingual group.

The factor of feature generalization is hypothesized to show an additive effect over multilingual

benefit in the acquisition of the non-native contrasts presented. Also, a shift in perceptual

assimilation types reflecting a trend towards learning is hypothesized in relation to the third goal

of the study. The appropriateness of the choice of language groups in order to examine the first

two goals has been discussed in previous chapters. The a priori possibilities from the three

language groups' comparisons can be made as follows. One or more of these states may emerge

as results of this study depending upon different hypotheses stated below.

Possibility (1) results if the presence of multilingual benefit emerges as dominant through

the perceptual performance of both bilingual groups, that is, Bengali-English (BE) and Spanish-

English (SE), will be significantly higher than that of the monolingual group, American English

subjects (AE). Possibility (2) reflects the scenario where only BE group may show higher









perceptual performance than either the SE or AE group, evidencing that only feature

generalization is facilitating the BE group to learn to perceive the non-native speech contrast

faster than the rest. There will be no evidence for multilingual benefit effect if the SE group fares

as equal to AE, the monolingual group in this case. Possibility (3) is concerned with the

assumption that feature generalization may be seen as an additive effect over the multilingual

benefit factor. In this case then, the BE group will perform significantly higher than the SE group

which in turn will perform higher than the AE group. In other words, the BE group will have

added advantage of enhancing the learning process since both multilingual benefit as well as

feature generalization factors will facilitate this process. Possibility (4) represents the fourth

logical alternative, that of null hypothesis. In this case, there will not be any significant changes

in the perceptual performances of all the three groups displaying no evidence for either

multilingual benefit or feature generalization.

Three different tests have been employed to investigate the hypotheses posited above. Data

of a total of sixty subjects, twenty from each language group, have been analyzed. The

perceptual performance of subjects in various tests like the identification test, AX discrimination

test and perceptual assimilation test has been measured by comparing the performance at the

pretest level to that at the posttest level. In order to normalize any performance differences at the

pretest level, difference in the scores of the posttest from the pretest have been submitted as data

to assess significance level.

The results presented in this chapter were tested for statistical significance using a number

of mixed-model ANOVA (Analysis of Variance) on various tests employed during pretest and

posttest phase of the experiment. The independent variables for the AX discrimination test,

identification tests (Pre-post ID task and Generalization task) as well as the perceptual









assimilation test remained same, such as native language group and contrast type. ANOVAs

reported the main effects of the independent variable, language group, as well as the interaction

between these variables, language group x contrast. In the discrimination test ANOVAs, the

dependent variable was mean percent correct score of a subject on sixteen trials of a particular

contrast type. In the varied identification test ANOVAs, the dependent variable was mean

percent correct score of a subject on twelve trials of a particular contrast. The percent change in

category assimilation type was used as the dependent variable for the perceptual assimilation

ANOVAs. If a particular fixed effect or their interaction was found to be significant at p<0.05 or

marginally significant such as p<0.1, then Tukey post hoc tests were conducted to examine the

differences of the least-squares means of the fixed effects in order to determine the statistical

significance of performance by individual language groups.

Results

This section gives a detailed account of the results in the areas of testing, such as

Identification task, AX discrimination task, generalization test and the perceptual assimilation

task. Since the results of perceptual assimilation require a more descriptive approach than a

statistical analysis, they are discussed in a separate sub-section below. Once the overall results

containing the main effect of independent variables are stated for the test of identification,

discrimination and generalization (provided immediately below), specific post hoc results are

discussed in detail under the headings for the relevant factors examined in this study.

Main Effects and Interaction of Independent Variables

In the case of the identification test on a trained-on talker, comparing the results of pretest

and posttest, no main effect for language groups was observed ( F(2,57)=2.42, p>0.05). Fig 3-1

shows the mean percent correct response in pretest and posttest level of identification test by

subjects from the three language groups. All three listener groups improved from pretest to









posttest, with the largest increases in performance observed with the Bengali-English group,

followed by the Spanish-English and American English trainees. The main effect for contrast

type was observed to be significant, as expected (F (3,171)=5.46, p<0.05). The interaction

between language groups and contrast was observed to be statistically significant

(F(6,171)=5.74, p<0.05).

100%
90% -
80% 74%
a 70% 63%
60% 56% EPretest
550%
protest and the posttest levPosttest
40%


20%
10%
0%
AE BE SE
Language Groups
AE=American English monolinguals; BE=Bengali-English
bilinguals; SE=Spanish-English bilinguals


Figure 3-1. Mean percent correct identification scores averaged over all four contrasts at the
pretest and the posttest level.

Since the language group contrast interaction was significant, post hoc tests using

Tukey's HSD procedure were conducted for only lateral, nasal and fricative contrast types. The

rhotics contrast showed no group differences. These results of individual contrasts are discussed

in post-hoc analysis under the section of individual factors.

The results of perceptual performance from the comparison of pretest and posttest level of

AX discrimination test show no main effect for language group ( F(2,57)=1.55, p>0.05, n.s).

Figure 3-2 represents the mean d' scores of AX discrimination test at the pretest and posttest









level for the three language groups. All three listener groups improved from pretest to posttest,

with the largest increases in performance observed with the Bengali-English group, followed by

the Spanish-English and American English trainees. The interaction between language group and

contrast was observed to be significant (F(6,171)=4.04, p<0.05). Of the post-hoc analysis using

Tukey test, only the lateral contrast showed statistical significance for language group effect.

3.00

2.50
1 4
Figure 3-2. Mean d' scores of AX discrimination test over all four contrasts at the pPretest
p1.50 1.30 Posttest








identification, showed no main effect for language groups (F(2,57)=.710, p>0.05, n.s). The


Figure 3-2. Mean d' scores of AX discrimination test over all four contrasts at the pretest and
posttest level (Maximum value of a d' score = 4).

In the case of the generalization test (an identification task), the scores were compared to scores

of the identification test (posttest) test in order to examine any evidence for the development

towards forming new phonetic categories for the non-native sounds by subjects in any of the

language groups. Posttest scores represented the performance level of subjects' perception of the

trained-on talker while the generalization scores displayed the accuracy in identifying the stimuli

from the novel talker (a voice not heard before during the experiment). The difference of the

posttest scores and generalization scores were submitted as data for the ANOVA test. The results

for the test of generalization, where similar stimuli from a new talker were tested for accurate

identification, showed no main effect for language groups ( F(2,57)=1.10, p>0.05, n.s). The









interaction between language group and contrast was observed to be statistically insignificant

(F(6,171)=1.21, p>0.05, n.s). None of the contrasts showed any significance concerning the

language groups. Therefore, no post hoc testing was conducted for the test of generalization.

Mean identification scores of the generalization test when compared with the posttest

scores, displayed unexpected results. Previous studies have provided evidence of development of

robust phonetic categories by L2 learners through generalization tests (Lively et al. 1994). Based

on their results, it was expected that subjects will perform either equally well or worse in

identifying the stimuli from a novel voice than the trained-on voice, as was the case in posttest

ID test. If any effects of multilingual benefit and/or feature generalization were to be observed,

then the expected results would be that either the BE (for feature generalization) or both the BE

and the SE groups (for multilingual benefit) would perform equally well, providing evidence for

the development of separate phonetic categories for the non-native contrasts. On the other hand,

the AE group, acting as the control group for both factors, would perform significantly worse in

the generalization test, displaying that AE group may require more training or exposure to these

contrasts in order to develop separate phonetic categories. The results from the current

experiment in terms of mean identification scores for the generalization test are provided in

Figure 3-3. For all three listener groups, the identification of the non-native contrasts was more

accurate in tokens from the new talker when compared with the trained-on talker, with the

slightly larger increase in performance observed with the Bengali-English group, followed by the

Spanish-English and American English trainees.










100%
90% 80% 82%
S80% 74% 75%
o 70% 63% 64%
60% Trained-on
S50% Talker
40%
oaNO Novel Talker
S30%
S20%
10%
0%
AE BE SE
Language Groups


Figure 3-3. Mean percent correct identification of contrasts spoken by trained-on talker
(posttest) and new talker (generalization test).

Therefore, looking at the overall results of the tests used to factor out any presence of

facilitating effects of multilingual benefit and/or feature generalization, it is difficult to conclude

that any of the hypotheses have been proven true. However, some of the post hoc analyses of the

ANOVA results, where marginal significance is observed in either main effect for language

groups or interaction between language groups and contrasts, are discussed in the following

section. These analyses may shed some light on the not-so-robust (perhaps subtle) effects of the

factors explored in this study.

Examining Evidence for Individual Factors

Multilingual benefit

As we have seen in the above section, the main effect for language in the pre-post

identification test was on the borderline of significance level (F (2,57)=2.42, p>0.05) (p=0.09).

The averaged scores over all four contrasts at the pretest and posttest level reflect this trend.

Looking at figure 3-4, we can see that both bilingual groups show a noticeable increase in the










accuracy scores from pretest to posttest level when compared with AE, the monolingual subjects'

acting as the control group.


100%
90% -
90%

70% -
80%



50% -
70%




30% -
60%
50%
40%
30%
20%
10%
0%


80%


63%


F~esT1
L~~2~esLI


Language Groups


100%
90% -
90%

70% -
80%



50% -
70%




30% -
60%
50%
40%
30%
20%
10%
0%


74%


63%


F~esT1
L~~2~esLI


Language Groups


Figure 3-4. Individual comparison of the monolingual group with BE bilingual group (top chart)
and SE bilingual group (bottom chart).

The mean percentage of improvement in the bilingual BE group is calculated as 19%

whereas the improvement percentage in monolingual AE group is observed as 12%. The mean

percent improvement in accuracy scores of the bilingual SE group is also seen as 19%, identical

to that of the bilingual BE group.









These differences of the posttest result from that of pretest results reflect the actual

perceptual performance by normalizing any initial performance differences at the pretest level.

Looking at the averaged improvement for all three groups, the prediction that both bilingual

groups individually have higher performance level than monolingual group appears to be

acceptable.

In addition to the above results, the interaction between language groups and contrasts was

significant (F (6,171)=5.74, p<0.05) (p=0.0001). Therefore, post hoc analysis was conducted

using Tukey procedure. Apart from the rhotic non-native contrast, all other contrasts lateralss,

nasals, fricatives) showed significance for the language group*contrast interaction. The lateral

non-native contrast was observed to be significant (F (2,171) =7.62, p<0.05). Similar

significance results were observed for the nasal non-native contrast (F (2,171)=4.54, p<0.05) and

the fricative non-native contrast (F (2,171)=3.75, p<0.05).

Of these, the only lateral non-native contrast was observed to show any evidence for the

facilitative effect of multilingual benefit in bilingual groups (BE and SE) since the performance

level of identification of the lateral contrast was significantly more among bilinguals than the

monolingual group. The nasal non-native contrast reflects a trend towards feature generalization.

Therefore the nasal along with the fricative contrast is discussed in detail under the next section.

Next, in order to determine the significant differences between each of the language group

pairs, the differences of least squares means were obtained using the post hoc Tukey test. When

analyzing the post hoc results for the lateral non-native contrast, it was observed that the

perceptual performance of accurately identifying the sounds in this non-native contrast by the

Bengali-English bilinguals was significantly better than the American English monolinguals

(p<0.01). The Spanish-English bilingual group's accurate perception of the lateral contrast was









marginally different than that of the American English language group (p = 0.09). Figure 3-5

shows the mean accuracy scores of identification of the lateral contrast at the pretest and posttest

levels by the three language groups. The mean percent improvement in performance by BE

group was the largest (25%), followed by SE group (18%) and AE group (8%). On the other

hand, when the differences of least squares means of the bilingual groups, BE and SE, were

compared using the Tukey post hoc test, the pair showed no statistical significance in their

performance from pretest to posttest level (p = n.s).

100% 90%
90%
S80% 74%
70%0 65%
60%i 55% 56% p Pretest
47%
th% i t o Posttest

C 30%
20%
10%
0%
AE BE SE
Language Groups



Figure 3-5. Mean percent correct response in identifying the lateral non-native contrast [1-[] in
the identification test at the pretest and posttest level.

These results for the lateral non-native contrast provide evidence for the facilitative effect

of multilingual benefit since Spanish speaking subjects, without any linguistic experience with

retroflex feature, were able to perform nearly as well as Bengali speaking subjects. The

monolingual language group (AE) was significantly (marginal with SE group) different from

both the bilingual groups.









Few of the post hoc testing for the AX discrimination test is suggestive of multilingual

benefit factor. As stated before, the main effect for language group in the improvement of

performance in case of discrimination test was not significant. However, the interaction between

language groups and contrasts was observed to be statistically significant. Only the lateral non-

native contrast was observed to have significant interaction (F (2,171)=9.86, p<0.05). The non-

native contrasts with manners as nasal (F (2,171)=0.46, p=n.s); fricative (F (2,171)=0.62, p=n.s);

as well as the rhotics (alveolar tap versus retroflex approximant) non-native contrast (F

(2,171)=0.04, p=n.s) were observed to have no significant interaction of language groups with

contrasts.

Analysis of the differences of least squares means in the post hoc Tukey test for the lateral

non-native contrast, showed that the improvement in performance by both bilingual groups (BE

and SE) individually was statistically significant over the performance by the monolingual group

(AE). Improvement in the Bengali speaking subjects' discrimination of the lateral non-native

contrast was more than the monolingual American English speakers (p < 0.001). The Spanish

speakers discriminated the lateral non-native contrast better than the monolingual American

English speakers (p < 0.05). However, the discriminability of the lateral contrast when compared

between the bilingual groups showed no significant difference (p = n.s). These results show that

the bilinguals were performing significantly better than the monolinguals in the case of lateral

contrast perhaps implying the presence of multilingual benefit effect. Figure 3-6 represents the

averaged d' scores of the discrimination test during pretest and posttest for the three subject

groups.










3.00

2.50
2.00
2.WPretest

1.50 Posttest

1.00

0.50

0.00
AE BE SE
Language Groups


Figure 3-6. Mean d' scores of discrimination of the lateral non-native contrast [1-[] at the pretest
and posttest levels by the subjects in the three language groups.

Looking at the post hoc analysis of the lateral non-native contrast in the identification test,

which was the same task as used during training, as well as the AX discrimination task provided

evidence for a trend that may direct towards demonstrating that among bilinguals, multilingual

benefit may perhaps be at work in enhancing the acquisition of the non-native contrasts more

than is observed in monolinguals.

Feature generalization

Feature generalization/feature productivity refers to the previous linguistic experience with

a native phonetic feature that may aid in the acquisition of a non-native contrast that exploits this

same feature but differs from any native contrast along other dimensions.

One of the hypothesis postulated as outcome for this study was that the Bengali-English

group may perform significantly better than the Spanish-English group due to its experience with

a retroflex feature in native contrasts. Additionally, both bilingual groups may perform

significantly better than the monolingual American English group. In this case, feature

generalization is expected to have an additive effect over multilingual benefit factor in this case.









It implies that when compared, within the bilingual groups, the Bengali-English group is

expected to perform significantly better than the Spanish-English group.

However, the results showed that feature generalization was not a significant factor. In

most of the ANOVAs and the Tukey post hoc analysis, no significant differences were observed

between BE and SE bilingual groups' improvement in perceptual performance. In the case of

Tukey post hoc analysis of the nasal non-native contrast during the identification test at the

pretest and the posttest level, the Bengali speaking subjects' improvement in performance was

significantly more than the American English speakers (p <0.05) but not significant from those

of the Spanish speaking subjects (p = n.s). The statistical significance was also not observed for

the Spanish-English bilinguals and the American English monolinguals (p = n.s).

100%
90% 80%
80%
68%
o 70%
51%






10%0
0% 0
AE BE SE
Language Groups


Figure 3-7. Mean percent identification scores for the nasal non-native contrast [n-nj by the
subject groups at the pretest and the posttest level.

Figure 3-7 shows the averaged identification scores for the nasal non-native contrast by the

language groups at the pretest and posttest level. All three listener groups improved from pretest

to posttest, with the largest increases in performance observed with the Bengali-English group,










followed by the Spanish-English and American English trainees. These results may suggest the

possibility that perhaps feature generalization may have been at work facilitating the Bengali

speakers to generalize the retroflex feature to the nasal non-native contrast.

One area where statistical significance is seen between Bengali-English and the Spanish-

English language group (p <0.05) is the post hoc analysis of the fricative palatall vs. retroflex)

non-native contrast within the pretest-posttest identification test ANOVA. Figure 3-8 provides

the averaged scores of identification of the fricative non-native contrast during the pretest and

posttest levels. All three listener groups improved from pretest to posttest, with the largest

increases in performance observed with the Spanish-English group, followed by the Bengali-

English and American English trainees.

100%
90% -
80% -
70%o 6% 67%
60 -57% 5 Pretest
60% 520
three language groups at the protest and the posttest level. Posttest
40%
30%
20%
10%
0%0
AE Be SE
Language Groups


Figure 3-8. Mean percent identification scores for the fricative non-native contrast [S-f] by the
three language groups at the pretest and the posttest level.

However, surprisingly in this analysis, both bilingual groups' performance in identifying

this particular contrast is similar to that of the monolingual group (p = n.s). This result does not

support the feature generalization hypothesis since the statistical significance between the BE









and the SE groups is showing inverse results (SE>BE). This result does not support the

multilingual benefit hypothesis also since BE bilinguals did not perform any better than the AE

monolinguals.

The interpretation of these results is discussed in chapter four. These results are suggestive

of the effect of multilingual benefit but not of the effect of feature generalization factor proposed

in the hypothesis. The feature generalization effect, expected to be seen in BE bilinguals because

of their linguistic experience with retroflex feature, appears to have no additive effect over the

general factor of multilingual benefit, expected in both the bilingual groups. Therefore, any

significant differences seen between the BE and the AE language groups can be suggestive of

multilingual benefit more so than of feature generalization.

The effect of multilingual benefit factor in enhancing the acquisition of non-native contrast

is observed exclusively in the non-native contrast with lateralization as the manner. The lateral

non-native contrast was considered the most difficult contrast for this study since all the three

language groups have only a single phonetic category for a lateral sound. Therefore, perhaps the

subtle presence of multilingual benefit and its effects seen in the bilingual groups' performance

surfaced in the acquisition of relatively difficult contrasts. The results seen with lateral non-

native contrast may reflect the positive influence that is not prominent enough to be observed

with easily discriminable contrasts. Perhaps by increasing the sample size of the language groups

may bring forth the factor of multilingual benefit as a robust effect. Therefore, to inspect for

more robust effects regarding the multilingual benefit factor, a reanalysis of the experimental

data was conducted. The reanalysis is discussed in the following section.

Re-Analysis

As seen in the results above, none of the tests or individual contrasts reflect any robust

effects of feature generalization. However, the presence of multilingual benefit is marginally









evident in the perceptual performance (both identification and discrimination tests) of the lateral

non-native contrast and as an overall result suggesting that the scenario of the bilingual groups

(BE and SE) performing better at identifying the non-native contrasts than the monolingual

group (AE) is accurate. These results, although not robust enough to be conclusive, provide us

with reasonable doubt and reflect towards a trend of the subtle presence of multilingual benefit

playing a role in enhancing the acquisition of non-native contrasts, particularly the difficult place

distinctions. Therefore, it is appropriate to further explore this area in a different light.

Since the results indicated improved perceptual performance by bilingual groups over the

monolingual group, it was thought suitable to examine the factor of bi/multilingualism as a

whole rather than two separate individual groups. Therefore, a re-analysis of the present data

(from the two tests provided at pretest and posttest levels: Pre-Post ID test and AX

discrimination test) was conducted.

Re-analysis: Multilingual Factor

In this re-analysis, data from both the bilingual groups, Bengali-English and Spanish-

English were put collectively under the bilingual group (N=40). The American English subject

groups was reiterated as the monolingual group (N=20). The pooled data was submitted to a

mixed-model ANOVA test where the independent variable, this time, was bilingualism vs

monolingualism, instead of individual language groups. The results of all the tests mentioned

above were re-examined with this ANOVA design.

Figure 3-9 represents the averaged accuracy scores for identifying all four non-native

contrasts at the pretest and the posttest level by the bilingual groups (BE and SE) and the

monolingual group (AE). The improved performance in percentage seen in subject groups after

training was 19% for the bilingual group and 12% for the monolingual group across all four

contrasts.










100%
90% -
80% 77%
70%M 63%
6558% EPretest5
60%

40%
30%-
S20%
10%
0%
Mono Biling
Language Groups



Figure 3-9. Mean percent scores of identification test across all four contrasts by the bilingual
groups (BE and SE) and the monolingual group (AE) at the pretest and the posttest
levels.

Interestingly, when pooled as a group, the bilingual subjects' perceptual performance in the

identification test was significantly better than the monolingual subjects' performance. The main

effect for multilingualism showed statistical significance (F (1,58)=4.75, p <0.05). The

interaction between the multilingualism and contrast was also significant (F (3,174)=4.45,

p<0.05). On examining the tests of effect slices, it was observed that of the four contrasts, the

lateral non-native contrast (F (1,174)= 11.72, p<0.05) and the nasal non-native contrast (F

(1,174)=5.03, p<0.05) showed statistical significance between the independent variable,

bilingualism-monolingualism.

In case of the AX discrimination test results, the pooled data set of bilinguals' performance

revealed that d' scores of bilinguals were marginally different from that of monolingual subjects.

The averaged d' scores of discrimination at the pretest and the posttest levels by the bilinguals

collectively compared to those of monolingual subjects has been represented in figure 3-10.The









main effect for multilingualism was borderline (F (1,58)=2.95, p = 0.09) unlike the insignificant

results (p = n.s) of discrimination test when examining individual language group effects.

3.00
2.50
r 2.00 1.79 --r e
1.50 1.30Posttest
1.06
1.00 0.85

< 0.50
0.00
Monolingual Bilingual
Language factor



Figure 3-10. Mean d' scores of AX discrimination test across all four contrasts, by the bilingual
groups (BE and SE) and the monolingual group (AE) at the pretest and the posttest
levels

The interaction between group and contrasts was observed to be significant

(F(3,174)=6.59, p<0.05). Therefore, post hoc analysis was conducted on individual contrasts

using the Tukey procedure. Of the four contrasts, only lateral non-native contrast showed

statistical significance between the bilingual and the monolingual groups (F (1,174)=18.50,

p<0.05). The post hoc results from the re-analysis were quite similar to the post hoc results from

the main analysis in terms of which contrasts showed significant differences among the

independent variables.

The results from the pooled dataset exhibit the effects of multilingual benefit being

employed by the bilinguals in order to enhance the acquisition of the non-native contrasts.

However, this re-analysis reveals the relevance of certain methodological issues. The effects of

certain factors like multilingual benefit perhaps cannot be captured at the standard sample size of









twenty subjects used in this study. These effects could be subtle and perhaps obscured by

individual subject variations. This limitation is further discussed in chapter four.

An Alternative Re-analysis: Only Place Contrasts

Alternative re-analysis of the data was conducted by examining individual contrasts used

for this study. The results from the main analysis and subsequently the results from the pooling

analysis reflected the need to investigate individual contrasts and to look for inconsistencies

among the four contrasts chosen for this study. The results that looked at individual language

groups showed enormous amount of variability in the means procedure of ANOVAs conducted.

In addition to that, the perceptual performances varied from one contrast to the other. Based on

the three reasons stated below, a re-examination was conducted of the four non-native contrasts

as hindsight.

Looking at the preliminary experiment, where data of initial AX discrimination of these

contrasts were gathered from native American English speakers, the highest d' score was

attained by the rhotics non-native contrast (d'=2.53) (refer to Chapter 2, pg 18, for details).

Though not at ceiling (Max d'=4), this particular contrast stood out as it showed ease of

discriminability at above chance (63%) whereas the other three contrasts showed discriminability

at chance level (50%). Therefore, ease of discriminibility of this contrast may have suppressed

the subtle effects that were seen in more difficult contrasts like laterals (prelim d' score = 1.13)

since the statistical analyses were conducted on the difference scores averaged over all contrasts.

This possibility gets support from the pretest scores for the rhotics contrast which were well

above chance level for all three groups (AE=59%, BE=67% and SE=67%). Thus, the probability

to observe subtle effects of multilingual benefit and/or feature generalization would have been

minimal.









Looking more closely at the phonetic features of this particular contrast, it appears to evade

the overall homogeneity of the stimuli for this study. The non-native contrasts with manners

lateral approximant, nasal and fricative only differ in terms of place distinctions. The lateral non-

native contrast has an alveolar and a retroflex [1-]. The nasal non-native contrast has an alveolar

and a retroflex [n-nj (see chapter Two for detailed discussion on dental/alveolar distributions in

Malayalam). The fricative non-native contrast has a palato-alveolar and a retroflex [J-g]

distinction. However, the rhotic non-native contrast [r-4] differs in not only the place distinction

but also the manner. This contrast involves an alveolar tap and a retroflex approximant.

Therefore, the rhotic contrast stands apart from the rest of the three only-place contrasts. Perhaps

this could have provided the contrast with inherent extra salience of acoustic features which may

have provided the subjects with ease of discriminability. This interpretation might raise questions

on the reasons for choosing rhotics contrast as one of the stimuli. In hindsight, initially it was

assumed that feature generalization might be a robust factor and will thus emerge even if there

were additional secondary cues to the contrast involving manner or other features that may

provide easier discrimination of the contrast. However, since this is not the case as is evident

from the main results, a reanalysis exclusively of the place contrasts is called for. One more point

of contention within this particular contrast is that the phonetic definition of retroflex

approximant sound in Malayalam has been a debatable issue among the phoneticians researching

this language (see chapter Two, pg 6, for details). In hindsight, no particular phonetic description

has been agreed upon making this particular contrast of concern regarding as stimuli for this

study.

Therefore, a second re-analysis of the data set, where only three non-native contrasts were

analyzed, was conducted. In this case, the scores on the non-native contrasts which differed only









in place distinctions (lateral, nasal, fricative) were used in a mixed-model ANOVAs for three

datasets: Identification test, AX discrimination test and the generalization test.

In the case of the identification tests at the pretest and posttest levels, the main effect for

language groups was observed to be significant (F (2,57)=5.12, p<0.05). Figure 3-11 shows the

averaged accuracy scores of identification test over the three place contrasts by all language

groups at the pretest and posttest levels. All three listener groups improved from pretest to

posttest, with the largest increases in performance observed with the Bengali-English group,

followed by the Spanish-English and American English trainees. The improvement percentage

for the bilingual groups was significantly higher. For the BE group it was 19% and for the SE

group it was 17%. The percentage of improvement in perceptual performance seen in AE group

was 9%.

100%
90% -
80% 77%
69%
S70%
60 57% 59% p Pretest
60% 52%
548%







AE BE SE






pretest and posttest levels.

The interaction of language groups with the three contrasts was also observed to be

significant (F (4,114)=5.81, p<0.05). The main effect of contrast showed no significant
significant (F (4,114)=5.81, p<0.05). The main effect of contrast showed no significant










difference, unlike the main analysis results, providing evidence for homogeneity within the three

place contrasts (F (2,114)=1.81, p=n.s.) (p=0.16).

The post hoc analysis using Tukey test was conducted. All three non-native contrasts

showed significant differences between the language groups. The post hoc results were the same

as the results from the main analysis since there was no change in the data set within the

individual contrasts.

In the case of AX discrimination test, the ANOVA results indicated marginal difference

between language groups. Figure 3-12 represents the averaged d' discriminibility scores over the

three non-native contrasts at the pretest and posttest levels. All three listener groups improved

from pretest to posttest, with the larger increases in performance observed with the Bengali-

English group, followed by the Spanish-English and American English trainees. The main effect

for language groups was seen at the borderline significance (F (2,57)=2.81, p=0.06). The

interaction between the language groups and the three place contrasts was observed to be

significant (F (4,114)=5.45, p<0.05).

3.00

8 2.50
0
2.00 17 Pretest

F 1.50 Posttest
S5 1.18
0.97
1.00 0.70 0.67

0.50

a 0.00
AE BE SE
Language Groups


Figure 3-12. Mean d' scores from the AX discrimination tests for the three place contrasts
(Laterals, nasals, fricatives) by the language groups at the pretest and posttest levels.









The results for the generalization test scores when compared with the posttest scores

showed no statistical significance, similar to the main analysis results. The area of testing for

robust phonetic category development did not reveal any significant differences among the

language groups in any of the analyses. One explanation could be the within group variability.

Individual subjects' performance may have dampened any differences between language groups.

Another explanation could be the issue of baseline talker intelligibility. These various

explanations will be discussed in detail in chapter four.

In summary, conducting re-analyses of the data demonstrated a different pattern of results

when 1) both bilingual groups were pooled as one group and 2) problematic contrast was

excluded. The positive influence of the multilingual benefit factor was supported with the

dependent measures when the issue of sample size was accounted for by pooling the BE and SE

groups together. The results for the identification tests showed significant group differences

between bilinguals and monolinguals and those of AX discrimination displayed marginal

significance in the performance of bilinguals over monolinguals. The reanalysis with only place

contrasts also revealed a significant main effect for language groups. However, effects of the

feature generalization factor did not surface prominently even in the reanalysis results. The next

section presents the results of the perceptual assimilation tests conducted at the pretest and

posttest levels during the training experiment.

Perceptual Assimilation Results

The perceptual assimilation task conducted at the pretest and posttest levels requires a

more descriptive interpretation of its results. Therefore, this section presents results solely from

these tests. Because of its descriptive nature, no prior predictions could be made on the results of

perceptual assimilation patterns. However, based on PAM-L2 predictions (Best & Tyler 2007),

we could expect to see a direction towards learning. For this study, learning was defined in terms









of a shift from non-categorical assimilation types like single category (SC) or UU (both

uncategorizable phonetic sounds) to assimilation types such as category goodness (CG) or

uncategorizable and categorizable type (UC) which reflect a trend towards higher sensitivity of

discrimination or even up to two-category assimilation which represents excellent discrimination

of the non-native contrasts.

The following sections describe results from the perceptual assimilation test in terms of

changes in assimilation types after training across all language groups. In addition, the question

of whether the effects of factors like multilingual benefit are seen in terms of perceptual

assimilation patterns will be explored by comparing the language groups' learnability trend

through the difference scores of assimilation types elicited from the pretest and posttest results.

Changes in Assimilation Types after Training

For this study, the assimilation types (for details, refer to chapter 1) described in PAM

model have been applied for interpreting the results. Five major assimilation types which reflect

various levels of assimilation, that is, two category (TC), single category (SC), category

goodness (CG), uncategorizable-categorizable (UC) and both uncategorizable (UU) have been

used. The results from the perceptual assimilation test reveal a consistent shift of assimilation

types from none / minimal separate categorization of non-native contrasts before training to

assimilation types like TC, CG and UC, which reveal improvement taking place in

discrimination of the non-native contrasts during training. This implies a direction towards

learning of the non-native contrasts as separate phonetic categories.

Figure 3-13 shows the mean percentage of number of assimilation types drawn from

subjects over all three language groups collectively. The results showed a decrease of 18% in SC

assimilation type and 5% in UU assimilation type from pre training to post training levels. The

TC and CG assimilation types displayed an increase of 11% and 10% respectively. The UC










assimilation type showed a modest increase of 2% from pretest to posttest. The overall change of

assimilation types from showing less categorization to a higher sensitivity towards categorization

provides evidence of learning taking place during the limited training period. The changes in

these assimilation types were not similar for all the four contrasts.

50% 46%
45% -
40% -
35% -
30% 8% 27%
25% 20%
20% 170 160/8% Pretest
15% -12% Posttest
10% -7%
5% -
0% r A
SC CG UU UC TC
Perceptual Assimilation Types

Figure 3-13. Mean percentage of assimilation elicited from all three language groups, pretest vs
posttest.

Looking at the results for the four non-native contrasts (figure 3-14) across all language

groups, decrease in SC assimilation type was observed for all contrasts. For the lateral and the

nasal non-native contrasts, there was found to be a steep decrease (-23% and -29% respectively)

when compared to the rhotic or the fricative contrasts (-12% and -10% respectively). This steep

increase is manifested as a large increase in CG (22% and 17% respectively) and TC (11% and

14% respectively) assimilation types for the nasal and lateral contrasts. This provides evidence of

possible learning of the non-native contrasts taking place. The rhotic contrast showed a modest

increase in CG (2%) but a comparable increase in TC (13%) assimilation types. In case of UU

(2%), it showed an increase unlike other non-native contrasts. The fricative contrast, on the other

hand, showed very modest increase in CG (2%) and TC (7%) types which were expected to










show learnability. Nevertheless, it displayed learnability through UC assimilation type where the

positive difference (improvement) is 12%.

30% -
22%
20% -7%
11143% 12%
10% -7.
4%
2% 2%2%
0 Lateral
0% Nasal
-10% -4% -5% 6%5% Rhotic
-10% 2-o % Fricative

-20% -

-23
-30%
-29%
TC UC UU CG SC
Perceptual Assimilation Type
Figure 3-14. Percentage of difference scores pretestt to posttest) of assimilation types for the
individual contrasts elicited from all three language groups.

On the whole, the SC and UU assimilation types displayed a decrease from pretest to

posttest whereas the CG and TC types showed an interesting increase. Overall, the subjects

showed more learning in lateral and nasal non-native contrasts than the other two contrasts.

Changes in Assimilation Types: Multilingual Benefit Factor

The perceptual assimilation test was one of the areas to examine the effects of factors like

multilingual benefit and feature generalization on the acquisition of non-native contrasts. These

chosen contrasts were phonemically assumed to be assimilated as the single category type at the

initial exposure pretestt) level. The assimilation results from the previous section provide

evidence of a definite shift towards learning these novel non-native contrasts in a very limited

training period. However, the question arises whether any language group differences were seen

pointing towards the effects of factors explored in this study.









The results from the identification tests and AX discrimination tests revealed no evidence

supporting the presence of feature generalization as a facilitating factor in the acquisition process

by the bilingual group, BE. Marginal effects were observed for multilingual benefit in case of the

lateral non-native contrast. On two alternate reanalysis of the data set, the presence of this factor

emerged as significant for the bilingual group when compared with monolingual group.

Therefore, it is reasonable to examine the bilingual groups as a whole comparable to

monolingual group rather than individual language groups.

Figure 3-15 displays percentage of assimilations elicited from individual language groups

at the pretest and posttest levels. The bilingual groups, BE and SE, showed similar shifts in

assimilation types from pretest to posttest, unlike the monolingual group. The SC type of

assimilation pattern was observed to have similar percentage at the pretest level for both BE

(38%) and SE (41%) groups, whereas the same assimilation type had a very high percentage of

60% at the pretest level for the monolingual AE group.

Interestingly, a sharp decline was seen in the SC type at the posttest level for both bilingual

groups, unlike the monolingual group. An inverse trend was observed for the TC and in some

cases for the CG assimilation type from pretest to posttest. Both bilingual groups showed a high

rise in the TC assimilation type at the posttest level contrary to the monolingual group. The SE

group was deviant from the BE group in case of UC assimilation type. The SE group displayed a

modest decrease of (-)5% whereas the other two groups showed similar percentage but increase

in UC type.

Taking a comparable view of the three groups, figure 3-16 shows the difference-scores of

the percentage of perceptual assimilation types comparing the two bilingual groups with the

monolingual group over all four contrasts. A positive percentage denotes an increase in the












American English Monolinguals 60%


18%
14%


0%

UC UU


Perceptual Assimilation Types


Bengali-English Bilinguals


14. 16%




TC UC UU CG
Perceptual Assimilation Types


Spanish-English Bilinguals


41%


20%
15%


25%


16%
11%


TC UC UU CG SC
Perceptual Assimilation Types


Figure 3-15. Percentages of assimilation types elicited from individual language groups at the
pretest and posttest levels are provided in three different charts.


* Pretest

* Posttest


70%
60%
50%
40%
30%
20%
10%
0%
-10%


* Pretest

* Posttest


70%
60%
50%
40%
30%
20%
10%
0%
-10%


* Pretest

* Posttest


I










percentage of elicited assimilation type from pretest to posttest and a negative percentage

denotes a decrease in the assimilation types gathered by the language groups.

25%

20% 16%
4 16%
15%
10%10%
6%
-25 14% 4%m
o 5%
0%

-5%-
I -40o 0 Be
w -5% _6-5%
S10% -6 Se

w -15%
'-14
-20% 18%

-25% -24%
TC UC UU CG SC

Perceptual Assimilation Types

Figure 3-16. Percentage of difference scores (posttest-pretest) of assimilation types elicited from
BE, SE bilingual groups and AE monolingual group over all four contrasts.

A steep decline was observed in the SC type along with a nearly equal steep increase in the

TC assimilation type, and CG type for SE group, during the posttest for bilinguals. On the other

hand, the monolingual group showed a decline in SC type lower in percentage than that of the

bilinguals. Moreover, the SC type remained high in percentage even at the posttest level for the

monolingual group. Only a modest increase was seen in the TC, UC and CG assimilation types

which reflected limited learning taking place within that group. Thus, the results described from

the perceptual assimilation tests provide evidence for the effects of multilingual benefit

facilitating the acquisition of novel non-native contrasts even with such limited training.

Although this section seemed to have required a more descriptive analysis of the

perceptual assimilation results, unlike the statistical results from other tests such as the









identification test and the discrimination test, a statistical analysis (one-way ANOVA) on

proportion change in assimilation types was conducted to examine significance based on the

individual language groups as well as the multilingual factor. The main effect for language was

observed to be significant (F (2,57)=4.16, p<0.05). In the post hoc Tukey test, BE group showed

proportionately more change in assimilation types than the AE group (p <0.05). The BE and SE

group showed no significant difference in proportion change (p = n.s). The SE group showed no

significant difference from AE group for proportion change (p = n.s).

These results lead to the reanalysis of the dataset by pooling the bilingual groups together

and comparing with the monolingual group. To look at the multilingual factor, a one-way

ANOVA was conducted which showed significant differences between the bilingual group and

the monolingual group (F (1,58)=7.78, p<0.05). These results, again, are suggestive of the effect

of multilingual benefit on acquiring the non-native contrasts by the bilinguals as a group,

regardless of their LI languages.

Summary

The results of the current study are suggestive of the effects of multilingual benefit. The

main language effects in reanalysis and post hoc tests of the lateral non-native contrast provided

support for the hypothesis stated for the multilingual benefit factor. The evidence for this factor

was mainly seen in the lateral non-native contrast in the dependent measures like the

identification test and the AX discrimination test. Both the reanalysis provided evidence for

multilingual benefit factor by showing significant group differences. However, neither the effect

of feature generalization nor its additive effect surfaced in any of the results. The only marginal

effect of feature generalization was seen in the identification test results for the nasal non-native

contrast, which does not show strong support for this factor. The test of generalization results

were observed as non-significant in terms of group differences as the mean percentage of









performance revealed that on an average subjects performed better with the novel voice

(generalization test) than with trained-on voice (posttest). The results from the perceptual

assimilation tests revealed a consistent shift towards learning of the non-native contrasts. Also,

they showed trends of effects of multilingual benefit among bilingual groups as opposed to the

monolingual group.

These results of this very limited training are only suggestive, of the hypothesis that

bilinguals can acquire non-native contrasts faster than monolinguals, given the facilitating effects

of multilingual benefit. However, no study is without its caveats. There were some unexpected

results such as subjects did better with novel voice than a trained-on voice. In addition, a lot of

within group variability was found for the generalization test. All the results, their interpretation

and caveats are discussed in the next chapter.









CHAPTER 4
DISCUSSION AND CONCLUSION

Introduction

The experiment was designed to address a number of important questions concerning third

language acquisition and cross-language speech perception: Would multilingual benefit and/or

feature generalization facilitate the acquisition of novel non-native contrasts among bilinguals?

Can perceptual assimilation patterns be modified through a limited training period? Would these

changes in assimilation patterns, if any, reflect differences between the bilingual and

monolingual subjects in terms of acquisition of non-native contrasts? These broad questions are

addressed through an overview of the results. In addition to that, a discussion regarding the

possible explanations for the unexpected results seen in this study will be put forth discussing

some unforeseen caveats in the methodology as well as issues with current assimilation models.

The results of this study will also be discussed in tandem with previous literature in order to

embed the results in a larger perspective, to further future potential hypotheses in the field of

language acquisition.

Overview of the Results

The design of the current study had three main aims. The first aim concerned the

assessment of the effects of metalinguistic ability developed as a result of having two languages

already in place. The second aim pertained to the assessment of the effects of previous linguistic

experience with certain phonetic features present in the target stimuli. The third aim was to

describe changes in assimilation patterns and examine any trends towards learnability. These

aims were narrowed down to more specific hypotheses. The choice of subject groups was

determined based on the formulation of the hypotheses. Briefly stated, Bengali-English (BE)

bilinguals were chosen because of their linguistic experience with the retroflex phonetic feature









and more so because of the fact that they were balanced bilinguals. Spanish-English (SE)

bilinguals were appropriate as the counteracting group solely because of their lack of any

linguistic experience with retroflex feature and also for having acquired two languages relatively

simultaneous. American English (AE) language speakers were chosen as the control group due to

lack of any substantial knowledge of another language, making them a monolingual group.

Based on these language groups, more strictly defined hypotheses concerning the main

aims were constructed. The effects of multilingual benefit would be evident if both bilingual

groups performed significantly better at identifying the non-native contrasts than the

monolingual group. The facilitating effects of feature generalization would be seen if the BE

group performed significantly better than the SE and the AE group. Additive effects of feature

generalization over multilingual benefit would be observed if the BE group identified and

discriminated the non-native contrasts more accurately than the SE group, which in turn

performed better than the AE monolingual group. Apart from these hypotheses, the results for

the generalization test were expected to be fairly equal to or lesser than the posttest identification

test results, since the successful generalization to a new voice would indicate the formation of

new phonetic categories corresponding to the non-native contrast. This prediction was based on

the results from the previous training studies examining generalization of newly acquired

contrasts to new voices (Logan et al 1991, 1993, Lively et al. 1994, Bradlow et al. 1997).

The results of the current study do not show robust empirical evidence to support the

hypotheses. However, this may be due to the complex nature of the stimuli (4 pair of novel non-

native contrasts to acquire) as well as very limited time period for implicit perceptual training (6

sessions- 35 minutes each). The subjects were required to perceive distinctions among eight

target sounds, four of which were novel for the subjects [1-, n-r- f-1b J -f]. Acquiring all these









four contrasts within the training period of only 3.5 hours (6 sessions of 35 minutes each) would

have been a very daunting task. Perhaps, over a longer period of time, the robustness of the

results will be seen. Therefore, to observe any robust effects in such limited training, the

influence of the factors being explored would have had to be really strong.

The hypothesis for multilingual benefit was supported in one particular contrast (lateral

non-native contrast) out of the four contrasts used as stimuli. The evidence for multilingual

benefit was consistent in the lateral contrast in the discrimination results as well as in all the

subsequent reanalysis conducted. The results of the perceptual assimilation test revealed effects

of multilingual benefit prevalent in the bilingual groups. The statistical analysis of the proportion

of change in assimilation patterns among the language groups was observed to be significant,

with the BE group having a larger change in assimilation patterns after training as compared to

the monolingual AE group but not significantly different from the bilingual SE group. The SE

group revealed a marginal significant difference in the proportion change from that of AE group.

Therefore, although the evidence supporting the effect of multilingual benefit in facilitating the

acquisition of non-native contrasts in bilinguals is not robust enough, it is suggestive of presence

of this factor.

The hypothesis of feature generalization/feature productivity did not surface as a

facilitating effect where multilingual benefit was not seen, such as a scenario where the BE

group would perform better than the SE and AE groups with no group differences between SE

and AE. The only instance which could be interpreted as perhaps an effect of feature

generalization was the results of the nasal non-native contrast in the case of the identification

test. However, even in this case, the SE group was at borderline significance with the AE group,

resulting in a very weak interpretation of the feature generalization hypothesis. In addition to









that, an additive effect that was hypothesized to result from having previous linguistic experience

with the retroflex feature in addition to multilingual benefit was not observed in any area of

testing.

The perceptual assimilation test results along with requiring a descriptive approach allow

us to make a priori prediction concerning the multilingual factor. It was expected that the

changes in assimilation types before and after training, if any, would reveal effects of

multilingual benefit; that is, both bilingual groups would show a greater shift towards the

assimilation types that clearly reflect learning of the non-native contrasts than the monolingual

group. The results of the study show support for this prediction. Group differences were

observed between the bilingual groups when compared with the monolingual group, where the

bilingual groups show a considerable decrease in SC (single category) and UU (uncategoriable)

assimilation types and an increase in the TC (two-category) and CG (category goodness)

assimilation types. This indicates that learning took place even during the very limited time

period of training. The learning was seen more in bilinguals than monolinguals.

An overall shift in assimilation types from within category non-discriminable types to

more discriminable categorical assimilation types were observed across all language groups and

all four contrasts. The PAM-L2 model (Best & Tyler 2007) provides us with an approachable

system of assimilation types that can be used to examine the perceptual assimilation performance

before and after training. However, some caveats in the model restrict us from further exploring

the direction of learnability through these specific assimilation types. These are discussed in the

next section in detail along with other unexpected results.

Alternate Explanations

If the results of the current study show support for certain hypotheses, there are other

results and cross-language group differences that are not so easily explained. This section









provides either possible explanations or interpretations of the various unexpected results found in

the study.

Test of Generalization Results

The results in the area of test of generalization during the posttest phase of the training

were contrary to the prediction. It was expected that the posttest scores of identification would be

higher than or fairly equal to the generalization test scores, as seen in previous literature.

Moreover, if the explored factors were in fact effective in enhancing the acquisition of the non-

native contrasts, it was expected that the bilingual groups would show near equal test scores for

posttest identification (trained-on voice) and generalization (novel voice) whereas the

monolingual group would show a lower score in generalizing to a novel talker than scores

attained with the trained-on talker. However, the results of this study supported no such claims.

On the contrary, there were no group differences between all three language groups. Moreover,

all three language groups displayed generalization scores (novel talker) that were higher than the

posttest scores (trained-on talker). On analyzing the individual contrasts in order to tease out

perhaps one particular contrast that may have been the easiest to discriminate, it was found that

no one contrast was consistently being identified accurately by all speakers. There was a

considerable variation in identification scores within groups and across contrasts. Therefore, one

particular contrast could not be held responsible for these unexpected results.

Another explanation for these surprising results could be the issue of baseline talker

intelligibility issue. This inherent/inbuilt intelligibility of an individual talker is perhaps the

factor affecting the results of the generalization test. Perhaps, the hyper-articulated sounds

produced by the novel talker made it easy for the all the subjects, across the language groups, to

identify the stimuli more accurately than was the case with trained-on talker. The term talker

intelligibility can be explained in regards to the articulation of the sounds by individual talker. It









could be seen as a continuum where on one extreme a talker produces "clear" hyper-articulated

speech which enhances the intelligibility of the speech for the listener of the same language. On

the other end of the continuum, a talker produces "plain" speech which may be unintelligible to

the listener, with same phonetic category inventory, in the context of various kinds of noises. The

terms "clear" refer to speech that is distinct and a part of speech production where the

acoustic/articulatory features are indicated clearly so as to be intelligible in adverse listening

conditions. It may perhaps allow the talker to enhance the distinctions between contrasts in a

phonological space (Uchanski, 2000; Smiljanic and Bradlow, 2005, 2007). The "plain" speech

refers to the normal rate of speech that a talker will produce in ideal listening conditions, that is,

not hyper-articulate the sounds in order to make one's speech intelligible.

Therefore, based on this interpretation, the results of the generalization test scores can be

explained in the following way. Probably, the novel talker used in the test of generalization

provided the stimuli, of which certain sounds were clearly hyper-articulated among the four

speech contrasts, unlike the stimuli from the trained-on talker used in the posttest. The hyper-

articulated sounds may have provided all the subjects with easy identification of those sounds

resulting in better accuracy than with the trained-on voice. With introduction to multiple talkers

during training, the formation of new phonetic categories, which was the desired effect, to

accommodate the distinctions between the new non-native contrasts may have started among the

subjects of the bilingual group. Therefore, with the baseline talker intelligibility varying

considerably for the trained-on talker and the novel talker, the subjects were able to perform

slightly better with the new voice than the trained-on voice. Although, no significant outcomes

were observed for this test, the results are still suggestive of the fact that a direction towards

development of robust phonetic categories corresponding to the non-native contrasts could be









seen since all groups fared better in generalization scores. The monolingual groups scored

equally well as they did with the trained-on voice, like the bilingual groups. Contrary to

hypothesis, it could be interpreted as an effect of the enhanced intelligibility of the novel talker

in increasing the performance level for the monolingual group. This interpretation of the results

also reflects the robustness of the training program already observed in many high variability

training studies (Logan et al., 1991; Lively et al., 1993, 1994; Bradlow et al., 1997).

Many perceptual training studies where multiple talkers are introduced during training

acknowledge the variations found in individual talker's intelligibility by the listeners as a

confounding factor for the extreme variations in the results (Lively et. al., 1993, 1994; Iverson et

al., 2005; Semiljanic and Bradlow, 2007 etc.). However, this area of inherent talker intelligibility

requires a deeper probe than a mere mention in the training studies as a confounding variable.

This aspect of perceptual training method can be looked at exclusively in future research work

which can contribute substantially in enhancing the efficacy and accuracy of the high variability

perceptual training method.

Necessity for Reanalysis

The design of the experiment was constructed with the assumption that the factors being

explored would show robust effects within the individual bilingual groups. Therefore, a standard

sample size of twenty subjects in each group seemed quite appropriate for the design.

Additionally, having a number of contrasts in the stimuli was expected to provide the study with

a range of perceptual data that could potentially contribute to the existing pool of explored

contrasts for the cross-language speech perception field and could be used for future work.

On the contrary, the outcome of the existing design showed no effects of the feature

generalization factor. Additionally, the effects of multilingual benefit showed only trends

towards its presence, and only in certain contrasts. No strongly suggestive support emerged









through any of the testing measures for these factors that may influence the acquisition of the

non-native contrasts among bilinguals when compared with monolinguals. A further inspection

seemed appropriate in order to examine the reasons for these varied unexpected results.

Therefore, two reanalysis were conducted looking at two issues: inefficacy of the sample size

and individual examination of the non-native contrasts.

Since the presence of the multilingual benefit factor was considered as merely suggestive

within each bilingual group in the main results of the study, looking exclusively at the

multilingual factor was considered as the next step. Combination of the two bilingual groups and

comparison with the monolingual group revealed significant group differences implying that

bilinguals had performed better at acquiring the non-native contrasts than the monolingual

groups. The pooling of bilinguals into one group provided evidence for multilingual benefit as

the bilingual group showed significantly better identification results than the monolingual group.

The reanalysis by pooling the two groups together displayed a significant effect of multilingual

benefit that gave the individuals with two or more language systems an ability to extract relevant

acoustic cues for the place distinctions of the non-native contrasts and be able to learn to identify

the sounds in these contrasts separately. However, this reanalysis revealed some caveats in the

original design of the experiment, namely, the issue of sampling size. It appears that by

increasing the sample size of the language groups, the effects of multilingual benefit become

robust. Additionally, a large degree of individual variation within the language groups pointed

towards the need for a larger sample size. The individual variability could have resulted from the

uncontrolled factors such as the individual aptitude for learning. The power issue in this study

brought forth the fact that multilingual benefit effects are not dominant enough to be seen as a

robust effect within a small sample of the population. In order to see significant group









differences brought forth with multilingual benefit as the determining factor, a larger sampling

size may be required, especially with very limited laboratory training period as in the current

study.

The other basis for an alternate reanalysis was examining all the non-native contrasts

individually. The reason behind this examination was to tease out any particular contrast whose

results may be suppressing an effect of multilingual benefit. The large variation in the perceptual

performance on these contrasts pre- and post-training was another motivating factor to conduct a

reanalysis.

Upon a closer inspection of the four contrasts in question, the non-native contrast that

differed in both place and manner appeared to be most different from the rest of the stimuli.

There were several reasons that emerged to exclude the rhotic non-native contrast from the

dataset and reanalyze it. The reasons have been discussed in detail in chapter three and will be

briefly restated here. Firstly, the results from the preliminary experiment showed that of the four

non-native contrasts, rhotics was the most easily discriminable by the monolingual American

English group. Therefore, we see high perceptual performance scores for the rhotics contrast.

Therefore, the results of this contrast may have suppressed the group difference significance

noticeably. Secondly, the sounds [r-4] of the rhotics contrast are distinguished in place as well as

manner unlike the other three contrasts the sounds of which differ only in place distinction. This

criterion perhaps provided the listeners with more of the salient acoustic cues resulting in easy

identification and consequently easier learning of this non-native contrast, which may have again

been the reason for not showing significant group differences in the main result. Moreover, the

phonetic definition of the Malayalam retroflex approximant has not reached consensus among

the phoneticians who have examined the Malayalam consonant inventory. Lastly, the fact that









Spanish-English bilinguals will already have separate established phonetic categories for the

alveolar tap, acquired through Spanish, and the alveolar approximant, acquired through

American English, making this contrast the most easily discriminable for the Spanish-English

bilinguals. This would not have been the case with Bengali-English bilinguals since they used

the Indian English dialect which substitutes the alveolar tap with the American English

approximant leaving the BE bilinguals as the only language group with no added advantage from

this contrast.

Together, these reasons motivated the exclusion of the rhotic contrast from the analysis.

Therefore, a reanalysis of the data set with only the place contrasts was conducted. The results of

the only-place contrast reanalysis, excluding the rhotic contrast results, for the identification test

revealed significant group differences among the three language groups, as expected. On post

hoc testing, the results of the lateral non-native contrast were suggestive of the effects of

multilingual benefit. The bilingual groups performed equally well at identifying the contrast. The

BE group performed significantly better than the AE group with SE group showing marginal

significance with the AE group. The results of AX discrimination test which had originally

showed no significant group differences, revealed a marginal main effect for language groups.

The results from this reanalysis were suggestive of the assumption that perhaps the results of the

rhotic non-native contrast were preventing the effects of multilingual benefit to surface as a

robust effect.

The results also shed light on the phenomenon that factors like multilingual benefit, which

enhance the perceptual performance or facilitate in the learnability of non-native contrasts, may

not be evident in a laboratory research with average difficulty contrasts, such as rhotics in this

study. In terms of inherent acoustic difficulty in perceiving the relevant cues of the contrast,









studies have shown that spectral properties are less easily discriminable than the temporal

properties when they are the distinguishing acoustic cue within the contrast. Alternatively, in

terms of previous linguistic experience/ native perceptual sound inventory, the PAM model of

cross-language speech perception (Best & Tyler, 2007) details that non-native contrasts which

are assimilated to a single phonetic category as equally good exemplars or equally bad exemplars

(SC type) or UU (no assimilation with less phonetic distance between them) type, will have a

higher level of difficulty in discriminability. According to PAM, during the learning process as

well, the learner will exert more time in overcoming the difficulty level of this contrast and

forming separate phonetic categories corresponding to the contrast. Therefore, only when the

target new non-native contrast is inherently at a higher level of difficulty will the effects of

multilingual benefit emerge as robust. Thus, conducting reanalysis of the dataset gathered from

this experiment revealed different perspectives of the factors explored in this study, which could

be potential candidate hypothesis for future work.

Caveats in Perceptual Assimilation Analysis

The results of the perceptual assimilation test were descriptive in nature. Reiterating the

results, an overall shift in assimilation patterns was observed across the three language groups

and all four non-native contrasts. A shift from assimilation types like SC (single category) or UU

uncategorizablee) towards more categorical/ higher sensitivity for discrimination assimilation

types like TC (two-category), CG (category goodness) or UC (uncategorizable-categorizable)

were observed. Noticeably, these results reflect the growth of the learning curve for all the three

language groups since a decrease in SC assimilation types was consistently observed from

pretest to posttest. Similarly, an increase was observed in assimilation types like TC, CG and in

some cases UC from pretest to posttest.









Although these learning shifts were seen for all of the language groups, a difference in the

percentage of shift was noticeable between the monolingual group versus the bilingual groups.

The bilingual groups, however, fared equally well, and their shifts in assimilation types were

comparable. Therefore, in order to examine group differences statistically, an analysis of the

proportion of change in the assimilation patterns was conducted which revealed significant main

effect for language groups as well as significant effect for the multilingual factor (during

reanalysis). The post hoc tests provided suggestive evidence for the multilingual benefit factor.

However, this analysis of proportion change concerns only the comparison of proportion of

change in the assimilation type which in no way describes the direction of learning. In order to

examine learnability among the groups in terms of perceptual assimilation, a definitive ranking

of the assimilation types, needs to be established. For this study, five assimilation patterns (SC,

UU, UC, CG and TC) were used to assess the perceptual performance of the learner groups.

These assimilation types were adopted from the revised PAM model (Best & Tyler 2007). Since

the current model of perceptual assimilation does not delve into the prediction of learning stages

based on the assimilation patterns revealed in the perception of an experienced learner before and

after training, the hypothesized learning continuum is proposed as an extension of the PAM

model.

To assess the direction of learning in the perceptual assimilation mode a continuum of

learning is postulated where SC (single category) and TC (Two-category) assimilation type form

the two extremes. According to PAM, SC assimilation type at the initial exposure reflects poor

discrimination of the given contrast since both sounds fall under one phonological category of

the existing language of the learner. The model predicts that the learner may not form a new

phonological category at all where SC assimilation takes place. In case of TC (two-category)









assimilation at the initial stage of perception, the type reflects the learner's good discrimination

of the contrast. The learner already distinguishes the sounds of the contrast into two different

phonological categories. Therefore, after becoming an experienced learner (in a classroom

setting or natural environment) or at the post-training stage (in a controlled setting), as is the case

in this study, if the learner maintains (pretest-to-posttest) the SC type of assimilation for a

particular contrast, no learning has taken place. In case the learner shows TC assimilation before

training and maintains the same level of assimilation post training, again no learning has taken

place since the learner is already maintaining a good distinction within the contrast. However, if

the learner progresses from SC type at the initial exposure stage to TC type at the post-training

stage, evidence of successful learning can be inferred. This shift in assimilation pattern would

provide evidence that the learner has been able to place the sounds into two different phonetic

categories at the end of the training. In general, any shift in assimilation patterns towards TC

during the course of training would evidence learning. Consequently taking these assimilation

types as the two extreme ends, a learning continuum is proposed (Figure 4-1).

Other assimilation types that are placed in this continuum are CG (Category-goodness),

UC (Uncategorizable-categorizable) and UU (both uncategorizable). The ranking of these

assimilation types can be speculated as follows. According to PAM, CG is a within-category

assimilation, like SC type, and allows poor to good discrimination. UU and UC are both

between-category assimilation patterns, similar to TC type, and may show moderate to excellent

discrimination. Therefore, to determine direction of learning, a ranking similar to this can be

assumed SC
shows shift in assimilation pattern from SC to CG or for that matter from any lower rank to a

higher rank (SC to UC, CG to TC or CG to UC), learning is taking place.









However, this ranking sequence is questionable. The ranking of SC and TC assimilation

patterns appear to be decisive on the continuum but this is not the case with CG, UU or UC.

Looking back at the PAM model for a more detailed description of these assimilation types, a

definitive distinction between the CG, UU and UC assimilation patterns seems missing.

According to PAM, CG assimilation type reflects the perception of a particular contrast as

being within the same LI phonological category but one phone being considered as the good

exemplar of the category and the other as being the deviant exemplar of the same category.

During the learning process, the learner may form a new phonetic category for the perceived

poor/deviant phone of the LI phonological category but only gradually. The good exemplar L2

phone may remain assimilated to the LI phonological category. This definition ranks the CG

type above SC type and lower to the TC type in terms of learning. Also, as a general assimilation

pattern, CG is a within-category type which restricts the placement of CG on the continuum

closer to SC as related to TC.

However, the assimilation type CG representing a within category assimilation pattern,

which may equate to non-learning type, may be seen with a different perspective. Based on the

results of the current study, it is observed that if the learner provides variable goodness ratings

(e.g. ratings 2 and 5 on the scale of 1 to 7) for the target contrast implying that he perceives the

sounds as good and bad exemplar of the same phonetic category at the pre-training level, and

over the course of training shifts to perceiving the same target contrast consistently as two

extremes of the ratings scale (at post-training ratings 1 and 7 on the scale of 1 to 7), then perhaps

successful learning does take place. However, according to PAM, the general assimilation

pattern of CG type is still within-category, that is, the learner still labels the modal response for

each sound of the contrast as the same. Yet the learner consistently perceives these sounds as









different (one sound at the goodness rating of 1 (highly dissimilar to the modal response chosen

from their native sound inventory) and the other sound at the goodness rating of 7(highly similar

to the modal response)) at the post-training level. The problem here lies with the orthography of

the language. Since this study was a controlled laboratory training experiment and learners were

asked to use an open set from their language systems for perceptual assimilation task, the

learners did not have new orthographic representations for the newly formed phonetic categories

to substitute and thus represent a TC type rather than a consistent CG type. However, in real life

situation, this dilemma may not arise as a learner may use the orthographic representations

corresponding to the phonetic categories of the target language while acquiring the non-native

contrast. In other scenarios, where orthography of the target language is not available or not

being used, the learner will have lexical items through which the phonemic distinction will be

clear and thus, the learner can eventually form separate phonetic categories for these sounds and

assimilate the contrast to a TC type. This may not be the case in controlled laboratory training

settings with a limited time period which may result in subjects still using the same modal

response but consistently discriminating the non-native contrast. Therefore, in this light, CG

types do show a trend towards learning.

Another case of perceptual assimilation is that an assimilation pattern is termed UC if only

one of the L2 phones is perceptually assimilated with an LI phonological category and

considered a good exemplar of that category. According to PAM, the discrimination of the

speech contrasts would be excellent as the learner would always perceive the fully assimilated

L2 phone separate from other less assimilated L2 phones. This places the UC type closer to the

TC type on the continuum which means that having UC type at post-training stage would show a

trend towards learning. However, with PAM specified definitions, the ranking of UC type to CG









type cannot be ascertained since CG type can allow for a wide range from poor to good

discrimination of the L2 contrast. Moreover, the UC assimilation type is not limited to just one

scenario as explained in the PAM model. Based on the results from this study it has been

observed that the learner provides various responses for categorization of the other L2 phone (the

uncategorizable one), one of which may overlap with the categorization of the fully assimilated

L2 phone. The discrimination, then, may not be as clear with the minimally contrasting words in

the target language. For example, the learner may assimilate one L2 phone consistently with LI

phonological category 'x' but assimilate the more deviant L2 phone to various other LI

phonological categories 'y', 'z' including 'x' category. In this scenario, UC assimilation type

cannot be interpreted to successful learning since at least ten to fifteen percent of the times the

L1 phonological category with which one L2 phone is fully assimilated, overlaps with the other

uncategorizable L2 phone. Therefore, even though the learner consistently identifies one L2

phone with one LI category, the identification of the minimal contrast of the sounds is not learnt.

Perhaps, in the real life situation, when the uncategorizable L2 phone is placed in minimal

contrast with other sounds of L2, the learner may gradually form a new phonological category

for it.

Similarly, the case of UU assimilation type is also quite ambiguous when it comes to the

placement on the learning curve. The UU type represents assimilation where the learner does not

assimilate the L2 phones to any one of his LI phonological categories. PAM claims that the

discrimination of this kind of speech contrast may depend on not just the comparative features of

the given sound to its closest LI counterpart but also its relationship to other phones in the same

phonological space. If the L2 phones are assimilated to similar set of LI sounds then

discrimination of this contrast is difficult for the learner since the acoustic proximity of the L2









phones to each other and to the set of assimilated LI phones will be less. On the other hand, if

the L2 phones are assimilated to different set of LI sounds the acoustic distance between the L2

sounds and the set of assimilated LI sounds will be much more and therefore would result in

easier discrimination by the learner. Thus in the case of UU type, the level of discrimination

depends on the inherent acoustic distance between those two sounds as well as the LI categories

to which they've been assimilated. Again, the possibility of overlapping of LI categories among

the sets of sounds that are assimilated to the L2 contrast is imminent in the UU type. Moreover,

the listener may be providing more than one modal response for each category and still not be

consistent with the goodness ratings. All these scenarios make it difficult to rank the UU

assimilation type against the UC and CG types.

Following definitions specified in the PAM model, assimilation types like category

goodness (CG) and the categorical (UC and UU) patterns cannot be allocated a ranking on the

learning continuum decisively. If the general assimilation pattern is considered then CG (within-

category) has to be placed lower than UU and UC (between-category). On the other hand,

predictions of PAM allocate a very broad range of discrimination levels from poor to good for

both CG and UU. Also, both UU and UC can possibly have lot of scenarios with varying

assimilation results as discussed above. Therefore, ranking these highly sensitive assimilation

types on the learning continuum solely based on PAM definitions of the assimilation patterns is

problematic.

Based on the perceptual assimilation results in this study, an attempt is made to speculate a

learning continuum which shows the assimilation types representing various stages of learning

L2/L3 by adults. Figure 4-1 shows the hypothetical learning continuum with five different

assimilation types representing different types of learning.










Direction of Learning


SC UU? UC? CG? TC
Figure 4-1. Hypothetical representation of the learning continuum displaying the ambiguity of
ranking the UU, UC and CG assimilation types in order to assess the direction of
learning.

The SC and TC assimilation types form the two extreme ends of the continuum where SC

type at the post-training stage equates to no learning taken place and the TC type represents

successful learning. The ranking of the CG, UU and UC assimilation types towards learning

cannot be decided without more empirical evidence supporting the ranking. Therefore, based on

the results of this study, the ranking of these three assimilation types can only be speculated.

Between CG and UU types, even though CG assimilation is a within-category pattern its

goodness ratings of the deviant L2 phone exemplar reveal that the learner perceives the L2

phones apart. Also, the perceptual assimilation results show a drastic increase in CG assimilation

types with consistently large goodness rating differences when pre-training and post-training

level assimilations are compared. UU assimilation type is a between-category assimilation

pattern which implies that the L2 phones are assimilated to different set of LI phones. Other

between-category assimilation patterns like TC denote successful learning but same cannot be

applied for UU type. As discussed above, UU assimilation type can result from many scenarios

which may not confirm as evidence of learning. Therefore, CG can be ranked above UU

assimilation type.

UC assimilation type can also result from many scenarios as discussed above. However, in

this case, one L2 phone is fully assimilated to one of the LI phonetic category which results in

better discrimination of the L2 contrast. Moreover, a shift towards UC assimilation type at the









post-training stage signifies that the learner has perhaps progressed from not distinguishing the

L2 phones at all to being consistently assimilating one L2 phone to one LI phonetic category.

This indicates progression towards learning and thus places the UC type above UU type and

more towards the right extreme of the learning continuum, that is, the TC extreme. The overall

perceptual assimilation results also validate this ranking as they show a decrease in the UU

assimilation type and an increase in UC assimilation type at the post-training stage. Between UC

and CG assimilation types, no definitive ranking can be stated between these two assimilation

types since both the types show an increase at the post-training stage. Therefore, both these

assimilation types can be positioned anywhere between UU and TC.

Thus, the ranking of assimilation types representing the learning continuum can be

hypothesized as SC
current study. There is a need to examine these assimilation types over a learning period and

determine whether they represent a direction towards learning. Further theoretical and empirical

exploration of these assimilation types will contribute substantially to confirm the learning

continuum and extend the current theoretical models of cross-language speech perception to

models of speech perception in the due course of learning. It is essential to assess the efficacy of

assimilation types like UU and UC in determining the perceptual assimilation patterns especially

regarding those that are elicited after training or in real life situations from early learners where a

shift in assimilation patterns can be determined in order to evaluate the progress towards

learning.

Results in Relation To Previous Studies

This study was an attempt to answer the following questions Does being a bilingual

facilitate in the acquisition of a third language? If so, is this facilitation influenced by the effects

of multilingual benefit which embody a developed general ability in balanced bilinguals to be









more sensitive to distinctions in auditory-acoustic stimuli and being able to abstract the

linguistically relevant information to contrast the speech sounds and develop a new phonetic

category? Secondly, would effects of feature generalization be observed in performance of

learners who may already have linguistic experience with a phonetic feature which is present in

the new non-native contrast as well? Lastly, what kinds of changes are seen in the assimilation

patterns of non-native contrasts post training? The results of this study contributed information

as well as a new range of data set to the existing literature in cross-language speech perception.

Previous studies have examined many factors that contribute towards the perception of

second language (L2) sounds and, in some cases, the subsequent establishment of L2 phonetic

categories. Phonetic, phonemic and acoustic factors among others have been determined to

influence the learning of L2 sounds (Tees and Werker, 1984; Polka, 1991; Best, 1995; Flege,

1995). However, learning a third language, unlike learning a second language, may be influenced

by additional factors attributed to the presence of two language systems in a bilingual instead of

one language system in a monolingual (Cenoz & Valencia, 1994; Ardeo, 2000; Munoz, 2000;

Bild & Swain, 1989). This hypothesis was based on the studies in lexical processing and word

learning by bilingual and monolingual children showing that bilingual children are more adept at

doing metalinguistic tasks and developing cognitive benefits (Peal and Lambert, 1962; Malakoff,

1992; Klein, 1995; Sanz, 2000 etc.). Numerous other studies looking at cognitive processing at a

linguistic level have explored the concept of multilingual benefit in bilingual and multilingual

children as well as adults which emphasized the fact that bilinguals with near native proficiency

in both languages show evidence of multilingual benefit (Bialystok, 1992, 2001, 2004; see also

review in Jessner, 2006). However, this concept of meta-linguistic awareness is not yet fully

explored in the field of cross-language speech perception and the acquisition of new phonetic









categories by bilingual learners of an L3. The results of the present study extend the research of

metalinguistic ability to the field of cross-language speech perception.

The study provides suggestive evidence that the bilinguals function at a level different

from monolinguals in reorganizing the information gathered from the speech stream in order to

deal with the high-demand task of acquiring new non-native speech contrasts. The results

support the claims of previous studies on lexical processing with respect to the effects of

multilingual benefit seen in acquiring a third language by bilinguals. For instance, the study by

Klein (1995) showed that during lexicon acquisition of a language, multilinguals learned a higher

number of lexical items than monolinguals. Klein based the explanation for this phenomenon on

enhanced cognitive skills in multilinguals which helped them tease out the potential relevant data

for resetting the particular parameter for the new language. Similarly in the present study, the

results of the lateral non-native contrast in particular showed that bilinguals, with or without

previous experience with the retroflex feature, displayed better perceptual performance than the

monolingual group. The study reflects the effect of multilingual benefit which may be further

explored as future research in speech perception.

The other factor that was explored in this study, feature generalization, did not show any

effects. The exploration of this factor was based on the assumption that features, an important set

of information contained in the established phonetic category, play a crucial role in developing

new phonetic categories (Nosofsky, 1986, 1987; Kruschke, 1992; Jusczyk, 1989; Lively et al.,

1993; Francis and Nusbaum, 2002). Few studies have looked at the feature generalization

phenomenon in the field of cross-language speech perception. Polka (1992) and Harnsberger

(1998) found results contrary to the expected results in this study that feature generalization will

reveal better perceptual performance. Nevertheless, a study by McAllister, Flege & Piske (2002)









found in the data from a few subjects that experience with the duration feature did facilitate

native American English speakers' perception of Swedish vowels (phonemic lengthening), so

that native AE speakers performed better than native Spanish speakers, who had no experience

with duration feature. The difference between this study and others was that the subjects were

'experienced' learners of Swedish and used the target language often. Their study was suggestive

of the hypothesis of feature generalization. Therefore, one factor that could explain the previous

studies conflicting results is that these studies (Polka, 1991; Harnsberger, 1998) examined only

naive perceivers with speech perception experiment. The present study employed an extended

approach which provided experience to the listeners through training, the next step from the

previous studies where feature generalization was tested on only naive listeners. The robustness

of the high variability perceptual training program was assumed to be effective in providing

sufficient linguistic exposure to the learners in order to bring forth any presence of the

facilitating effects of these factors.

The present study was unable to provide the missing link between the studies where the

feature generalization factor was explored but not found and the study which indirectly

suggested the presence of feature generalization. Regardless of gaining from perceptual training

with multiple talkers and high variability in stimuli tokens, Bengali-English speakers were not

able to reach a native-like performance level. This language group's perceptual performance in

both testing areas of identification and discrimination was not indicative of any robust effects of

feature generalization. Although, Bengali-English speakers performed significantly better than

the monolingual group of American English speakers, they were unable to display an additive

effect over the metalinguistic ability found in the Spanish-English speakers as well. Since both

Bengali-English speakers and Spanish-English speakers performed near equally in most of the









testing measures, the performance of a bilingual group (SE) without any previous experience

with the retroflex feature negated any trace of the presence of feature generalization that may

have been possible. Thus, the present study confirmed the results of the earlier studies (Polka,

1992; Hamsberger, 1998; Tajima et al., 2008) that whether the listeners are naive or received

limited implicit training, they are unable to generalize the phonetic features with which they have

had previous linguistic experience. Additionally, recent studies have shown that only phonemic

experience/ language experience over time facilitates the perception of non-native contrast and

that simultaneous bilinguals acquire native like phonetic categories which are developed fully by

adulthood (Silver, 2007; Sundara et al., 2006, 2007). This may explain the results of the current

study where no feature generalization was observed. The lack of extensive exposure of the

stimuli to the language groups is one of the explanations provided for results showing no support

for the feature generalization hypothesis. This possibility can be explored in future research work

by elongating the span of training and providing extensive exposure of the non-native contrast to

subjects. Extensive training may result in observance of effects of language experience in this

respect.

The third focus of the study concerned the role of the bilinguals' particular perceptual

category inventory in the acquisition of non-native contrasts from a third language. Several

models have been developed to account for the influence of native perceptual categories on the

perception and acquisition of non-native speech sounds, including the Speech Learning Model

(SLM) and, of greater interest for this project, the Perceptual Assimilation Model (PAM). PAM

concerns the discriminability and learning of non-native contrasts based on the relationship

between the non-native and native category inventories. These relationships have been

encapsulated in several assimilation types which were adopted for this study. The results of the









present study assessed the extent to which initial patterns of assimilation change following

limited training experience and measured the effects of the multilingual factor within the

perceptual assimilation pattern shift. The results revealed a shift of perceptual assimilation types

towards perceptual learning. On an average, the assimilation types SC and UU decreased and the

assimilation types TC, CG and UC increased from the pre training phase to post training phase.

However, the relationship of learning could not be deciphered as the PAM model does not

provide a definitive ranking of UU, UC (categorical assimilations) and CG (within category

assimilation) assimilation types with respect to learning. In addition, the multilingual effects

were also observed within the results of perceptual assimilation tests. The results were suggestive

of the positive perceptual learning seen among bilinguals which appeared to be higher than the

meager amount of perceptual learning that took place among the monolingual group. This study

provides a substantial empirical dataset which can be used to explore the issue of learnability

ranking among the assimilation types as future research work.

Another area of future research that could be pursued would be to examine the rhotic

segment orthographicc symbol: Y') of Malayalam which lacks definitive phonetic description.

This sound, as discussed in Chapter 2, has been presented in previous literature with various

articulatory descriptions (Kumari 1972, Sreedhar 1972, Asher & Kumari 1997, Kalackel 1985,

Raja 1960, Bright 1998, Krishnamurti, 2003). A clear phonetic description of this particular

segment of Malayalam consonant inventory is required. One method that can be used to describe

the features of this rhotic segment is acoustic analysis. It is observed that most of the descriptions

affirm the feature of retroflexion in this sound. Previous studies on Tamil liquids also report the

occurrence of similar sound which includes features of retroflexion and central airflow

(Narayanan et al 1996, 1999, McDonough & Johnson 1997). This type of research of the









Malayalam sound as well could provide a definitive phonetic description of the sound which

could be compared to similar sound segment in other Dravidian languages and contribute to the

knowledge of phonetic sounds of the Malayalam language.

Conclusion

This dissertation extends the research on cross-language speech perception and language

acquisition to a new range of contrasts, different manners such as lateral, nasal, fricative and

rhotic with alveolar retroflex as the place distinction. The study is successful in determining

whether current theories and established claims of multilingual benefit in the field of lexical

processing were generalizable to the area of cross-language speech perception in regards to

learning new non-native contrasts similar to that of leaning of the lexicon. The results suggested

that the effects of multilingual benefit were prevalent in the learning of novel contrasts with

limited training period. However, the claim on the effects of feature generalization received no

support in the results, therefore, confirming the findings of the earlier studies.

In the area of cross-language perceptual classification, the study provides a range of

empirical datasets for assessing the cross-language differences in the perceptual assimilation

patterns observed before and after training. It points to open questions of ranking of assimilation

types in terms of learnability that may strengthen the predictability of perceptual assimilation in

determining the direction of learning. These concerns need to be addressed by the current models

of perceptual assimilation. Beyond these general findings, the study also demonstrated the need

for larger sampling size in order to get robust effects since the effects of multilingual benefit

observed were merely suggestive in nature and seen along with lot of individual variations.

Overall, the results presented here lead to suggest a productive line of research for future work

discussed in the above section. The caveats such as looking into the baseline talker intelligibility

issue within training and testing; establishing learnability ranking of the assimilation patterns;









examining with exclusive focus on the effects of multilingual benefit were revealed during the

progress of this study. These caveats will provide substantial insight and enhance the

productivity of the future research.

In summary, the results supported only the hypotheses of multilingual benefit postulated

in this study. The factor of multilingual benefit is brought forth and explored within the realm of

cross-language speech perception field through the results of the study.









APPENDIX A
PRELIMINARY SCREENING

Subject Information form



Subject code (not to be filled by participant):

Name:

Age:

Email:

First languagess:


Language(s) you speak with your parents/grandparents/siblings:



Spoken proficiency in languages other than first languagess:



Language classes that you might have taken in Middle/high school:



Countries that you've visited (also give length of stay if more than 2 months):





History of any hearing impairment:



If participating in the study for credits specify name of course and Instructor:









Screening for Spanish Spoken Proficiency

This was a brief spoken test conducted at the time of screening the monolingual subjects as

well as bilingual Spanish-English subjects. The subjects were asked to say out loud, at normal

speaking rate, two phrases which are considered as tongue twisters in Spanish. Then they were

asked whether they were aware of the meaning of the phrase or certain words in them. This brief

proficiency test was conducted to assess the level of spoken proficiency as well as lexical

knowledge of the Spanish language. In case of monolinguals, if the candidates were not able to

maintain a normal speaking rate with less than 40% correct pronunciation (that is, they

substituted English sounds like [a] instead of [r], [t] instead of [t], [g] instead of [y] and dropped

the [e] at the end of word "roque" etc.), they could safely be considered under the monolingual

language group, provided they had exposure to no other language. In case of self-reported

Bilingual candidates, high level of spoken proficiency was considered along with complete

knowledge of lexical items and sentence structure.



Hay tres tristes tigres en un trigal.
[ai t res tris tes. ti yre se nun tri yal] IPA transcription

El perro de san roque no tiene rabo.
[el pe. ro. e sdn ro ke no tie ne ra 3o] IPA transcription









First time interested participant email:


Thanks for the interest in the study!
You can see more details about the study at
http://web. csd.ufl. edu/langbrain/Other Research.htm#_Perception training_study_2

Which of the language groups would you put yourself into? Before I go ahead and recruit you,
you'll have to provide me with some information on your language background.

At what age did you start learning your second language, if any? What other languages have you
been exposed to and for how long? (Please specify middle/high school, college, trips to places
etc.)
Also, did you work in any school or college projects involving a different language? Please let
me know, if you've had any roommates or friends from whom you were exposed to a language
other than English.
What is your major, minor and year at UF? Also, where did you learn about the study?
If you've any questions, please feel free to ask!

Thank you!

Divya


129










APPENDIX B
SUBJECT INFORMATION

Individual Subject Attributes


Spanish-English Speakers


Name code Age Spanish Dialect First Language(s) Age of Exposure (L2) Linguistic background
Portuguese average proficiency, French -
SE01 20 Venezuelan English, Spanish -- r
low proficiency
SE02 23 Panamanian English, Spanish -- exposure to Arabic
SE03 20 Salvadorian Spanish 2yrs English, 2 years of French
SE04 21 Cuban Spanish, English -- N.A
SE05 23 Argentinean Spanish 4yrs English, Portuguese- average proficiency
SE06 19 Colombian Spanish 4yrs English, a little French
SE07 20 Colombian English 6yrs Spanish
SE08 20 Mexican Spanish, English -- Introductory French and Portuguese
SE09 19 Cuban Spanish 4yrs English
SE10 20 Panamanian English, Spanish -- N.A
SE11 22 Cuban English, Spanish -- N.A
SE12 19 Salvadorian Spanish, English -- N.A
SE13 20 Nicaraguan Spanish, English -- N.A
SE14 19 Colombian Spanish, English -- 2 years of French (Middle school)
SE15 19 Mexican English, Spanish -- a little French
SE16 20 Mexican Spanish 4yrs English
SE18 19 Cuban Spanish 3 yrs English
SE19 21 Mexican Spanish, English -- N.A
SE20 21 Colombian English 2yrs Spanish
SE21 20 Nicaraguan Spanish, English -- N.A










Bengali-English Speakers


Name Age of exposure
Code Age City, State First Language(s) (L2) Linguistic Background
Code (L2)


Kolkota, West Bengal
Kolkota, West Bengal
Kolkota, West Bengal
Kolkota, West Bengal
Kolkota, West Bengal
Kolkota, West Bengal
Kolkota, West Bengal
Kolkota, West Bengal
Kolkota, West Bengal
Kolkota, West Bengal
Kolkota, West Bengal


BE12 24 Kolkota, West Bengal

BE13 25 Kolkota, West Bengal
Chandan Nagar, West
BE14 27
Bengal
BE15 27 Chakda, West Bengal
Barddhaman, West
BE16 29
Bengal
BE17 28 Dhaka, Bangladesh
BE18 30 Kolkota, West Bengal
BE19 28 Kolkota, West Bengal
BE20 32 Kolkota, West Bengal


Bengali
Bengali
Bengali
Bengali, English
Bengali
Bengali
Bengali
Bengali, English
Bengali
Bengali
Bengali

Bengali, Hindi


Bengali

Bengali

Bengali

Bengali

Bengali
Bengali
Bengali
Bengali


8 yrs
6yrs
3yrs


3yrs
6yrs
8yrs


8yrs
2 yrs
8yrs


6yrs

8yrs

3yrs

5yrs

12yrs
3yrs
6yrs
8yrs


BE01
BEO2
BE02
BE03
BE04
BE05
BE06
BE07
BE08
BE09
BE10
BE11


English, Hindi
English, a little Hindi
English, formal education in Hindi
English, Hindi, Sanskrit
English, formal education in Hindi
English, a little Hindi
English, Hindi
English, Hindi, exposure to Assamese and Gujarati
English, a little Hindi
English, Hindi, exposure to Urdu and Kannada (3 months)
English, formal education in Hindi
English, formal education in Hindi, exposed to Telugu (6
months)
English, a little Hindi

English, Hindi, exposure to Kannada (20 months)

English, a little Hindi and Nepali

English, Hindi, Sanskrit, German, a little Urdu

English, a little Hindi, Assamese and Oriya
English, a little Hindi and Assamese
English, Assamese, a little Hindi
English, Hindi










American English Speakers


Name Code
AE01
AE02
AE03
AE04
AE05
AE07
AE08
AE09
AE10
AE11
AE12
AE13
AE14
AE15
AE16
AE17
AE18
AE20
AE21
AE22


First Language(s)
English
English
English
English
English
English
English
English
English
English
English
English
English
English
English
English
English
English
English
English


Linguistic Background
Spanish Middle/High school
2 years of Spanish
Spanish Middle/High school
a little Spanish
Introductory French and Spanish
a little Spanish and French
a little Spanish
2 years of Latin, exposed to Spanish
1 year of Latin
2 years of Spanish(High school), years of Latin (Middle School)
I year of Spanish (Middle School)
I year of Spanish (Middle School)
3 years of Spanish (Middle/High school)
3 years of Spanish (High school)
a little Spanish and Italian
a little French, exposure to Spanish
2 years of Spanish (Middle School)
2 years of Spanish (High School), I year of Latin
2 years of Spanish (Middle School)
I year of Spanish (Middle School)









APPENDIX C
EXPERIMENT INSTRUCTION SHEET

Instructions given to subjects on the first day of the experiment (Pretest Phase):

Session 1 Instructions

Welcome to the Linguistics Laboratory. We appreciate your participation in this

experiment and also hope you will find it interesting.

You will be participating in a study of the perception training of new sounds (consonants -

for example, the first sound "1" in the word "lamp") from another language. This training study

will carry on for 8 sessions. Today is your first session. In this session, you will have to do four

tasks.

Task 1 Familiarization:

In this task, you will listen to consonants from another language. Each consonant sound

will be preceded and followed by a vowel sound (example : "ana" or "ulu"). Each consonant

sound, regardless of the vowels, is associated with a symbol. You will hear a word like "ulu" and

then see a symbol 1" You do not have to click any response button for this task. All you've to

do is learn to associate the consonant sound with the symbol. This may be difficult initially as the

sounds and symbols are from another language. This will take no more than 6 minutes.

Task 2 Identification task:

In this test, you will listen to one word like "ana" and then you've to choose the best

possible symbol that you identify the consonant sound with (example: after you hear "ana", you

click on "n"). There will be total of 8 consonant symbols (1, L, n, N, sh, S, r, R) to respond from.

Again, use the knowledge gained from the familiarization task. Since some of the sounds will be

new to you, they may be difficult for you to tell apart. However, even if you're not entirely sure









which consonant symbol it is, just make the best choice possible and move on to the next sound.

This will take about 15-20 minutes.

Task 3 Discrimination task:

In this test, you will hear two words in a row, separated by a short pause. These words are

from another language, and each word consists of a consonant preceded and followed by a vowel

sound. Your task is to determine whether the two consonants are the same consonant or different

consonants. If the two words have the same consonant, click on the "same" button on the screen.

If the two words have different consonants, click on the "different" button on the screen.

For example, you might hear something like the following:

ala (pause) ana

In this example, the consonant in the first word is "1" and the consonant in the second word

is "n". Since these are different consonants, you would choose the "different" button. This is a

relatively easy example. However, in this test, you will be hearing words from another language.

Some of these words may be difficult for you to tell apart. However, even if you are not entirely

sure if the consonants are the same or different, just make the best choice possible and move on

to the next set of words.

Task 4 Instructions for Perceptual Assimilation task:

In this task, you will listen to words from another language, similar to previous tasks. You

must choose the best possible sound from your native language that closely corresponds to the

sound given. Write the letter (or letters) best representing the speech sound on the sheet

provided. Then compare the non-native sound with closely related sound from your own

language. On the scale from 1-7, circle the number that you feel is appropriate in comparing the

similarity of the non-native sound with that of your own language sound.









For Example, if you hear "ala" and your native language is American English, you may

choose the closest sound from you language as being "1 as in laugh". If you choose this, write "1"

on your answer sheet and then scale the similarity of the two sounds. How similar are they? If

they sound exactly same to you, give the scale number "7- exactly same". Similarly, you can use

other options on the scale as well depending on your judgment.

1 2 3 4 5 6 7

Very different exactly same

This is a paper-pencil task, so you'll listen to the sound via headphones and record all your

responses on the sheet provided.










APPENDIX D
PERCEPTUAL ASSIMILATION RESPONSE SHEET


Note: This is a sample page of the perceptual
were elicited in this task.

Name:
Serial Letter Goodness Rating (1-7)
No. symbol


assimilation response sheet. In all, 80 responses


Subject Code:


1
Very different

1
Very different

1
Very different

1
Very different

1
Very different

1
Very different

1
Very different

1
Very different

1
Very different

1
Very different

1
Very different

1
Very different

1
Very different


Task:


7
exactly same

7
exactly same

7
exactly same

7
exactly same

7
exactly same

7
exactly same

7
exactly same

7
exactly same

7
exactly same

7
exactly same

7
exactly same

7
exactly same

7
exactly same










APPENDIX E
RESULTS IN TABULAR FORMAT


Table E-1. Mean percent identification scores averaged over all four contrasts at the pretest and
the posttest level.
Language groups Pretest scores Posttest scores
AE 51 63
BE 61 80
SE 56 74

Table E-2. Mean d' scores of AX discrimination test averaged over all four contrasts at the
pretest and posttest level.
Pretest Posttest
Language groups scores scores
AE 0.85 1.3
BE 1.17 1.94
SE 0.95 1.64

Table E-3. Mean percent identification scores of contrasts spoken by trained-on talker (posttest)
and new talker (generalization test).
Language groups Posttest scores Generalization scores
AE 63 64
BE 80 82
SE 74 75



Table E-4. Mean percent scores of consonant identification test for all four contrasts by the
bilingual groups (BE and SE) and the monolingual group (AE) at the pretest and the posttest
levels.


Language groups


Laterals
Pretest Posttest
47 55
65 90
56 74


Nasals
Pretest Posttest
51 59
59 80
54 68


Fricatives
Pretest Posttest
47 57
52 62
46 67


Rhotics
Pretest Posttest
59 82
67 87
67 88


Table E-5. Mean d' scores of AX discrimination test across all four contrasts, by the bilingual
groups (BE and SE) and the monolingual group (AE) at the pretest and the posttest levels.
Language groups Laterals Nasals Fricatives Rhotics
Pretest Posttest Pretest Posttest Pretest Posttest Pretest Posttest
AE 1.00 1.04 0.47 0.76 0.63 1.10 1.31 2.29
BE 1.58 2.77 0.90 1.38 0.99 1.45 1.21 2.17
SE 0.86 1.75 0.49 0.73 0.66 1.05 1.81 3.04










Table E-6. Reanalysis results: Mean percent identification as well as AX discrimination scores
averaged over all four contrasts at the pretest and the posttest level for the bilingual groups (BE
and SE) combined versus the monolingual group (AE).
Language factor ID Test Score AX Discrimination score
Pretest Posttest Pretest Posttest
Monolingual 51 63 0.85 1.06
Bilingual 58 77 1.3 1.79

Table E-7. Mean percent correct response (averaged over contrasts) for identification of the
three place-contrasts lateralss, nasals, fricatives) by bilingual (BE and SE) and monolingual (AE)
groups at the pretest and posttest levels.
Language groups ID Test Score AX Discrimination score
Pretest Posttest Pretest Posttest
AE 48 57 0.7 0.97
BE 59 77 1.15 1.87
SE 52 69 0.67 1.18









LIST OF REFERENCES


Agresti, Alan & Finlay, Barbara. (1997). Statistical methods for social sciences (3rd Ed.).
Prentice Hall.

Aakhane-Yamada, R., Tohkura, Y., Bradlow, A.r., & Pisoni, D.B. (1996). Does training in
speech perception modify speech production? Spoken Language, 1996, ICSLP 96,
Proceedings., Fourth International conference on. Vol. 2, pp. 606-609.

Asher, R.E. & Kumari, T.C. (1997). Malayalam (Descriptive Grammars). London & New York:
Routledge, 1997. Pp. xxvi, 491.

Aiyar, L. V. R. (1972). "Two Malayalam phonemes," International Journal of Dravidian
Linguistics, 1(2) 10-18.

Best, C.T., McRoberts, G.W. & Sithole N.M. (1988). "Examination of perceptual reorganization
for nonnative speech contrasts: Zulu click discrimination by English-speaking adults and
infants," Journal of Experimental psychology: Human perception and Performance, 14,
(3), 345-60.

Best, C. T., and Strange, W. (1992). Effects of phonological and phonetic factors on cross-
language perception on approximants. J. Phonetics 20, 305-330.

Best, C. T. (1994). The emergence of native-language phonological influences in infants: a
perceptual assimilation model, in The Development of Speech Perception: The Transition
from Speech Sounds to Spoken Words, edited by H. C. Nusbaum ,MIT, Cambridge, MA.

Best, C. T. (1995). A direct realist perspective on cross-language speech perception, in Speech
Perception and Linguistic Experience: Theoretical and Methodological Issues in Cross-
language Speech Research, edited by W. Strange ,York, Timonium, MD, 167-200.

Best, C. T. and Tyler, M. D. (2007). Nonnative and second language speech perception:
Commonalities and complementarities in Language Experience in Second Language
Speech Learning: In Honor of James E. Flege edited by Ocke-Schwen Bohn and Murray J.
Munro, John Benjamins publishing.

Bhattacharya, Krishna. (1988). Bengali Phonetic Reader. Central Insitute of Indian Languages,
Mysore.

Bialystok, E. (1992). "Selective attention in bilingual processing," in Cognitive Processing in
Bilinguals. Edited by R. Harris, Amsterdam 501-514.

Bialystok, E. (2001). "Metalinguistic aspects of bilingual processing". Annual review of Applied
Linguistics, 21, 169-81.

Bialystok, E., Craik, F., Klein, R. & Viswanathan, M. (2004). "Bilingualism, aging and cognitive
control: evidence from the Simon task," Psychology andAging, 19 (2) 290-303.









Bild, E.R. and Swain, M. (1989). "Minority language students in a French immersion
programme: Their French proficiency," Journal of Multilingual and Multicultural
Development, 10, 255-274.

Bradlow, A.R., Pisoni, D.B., Aakhane-Yamada, R, & Tohkura,Y. (1997). "Training Japanese
listeners to identify English /r/ and /1/: IV some effects of perceptual learning on speech
production," J. Acoust. Soc. Am. 101 (4).

Bright, W. (1998). The Dravidian Scripts. In The Dravidian Languages, edited by Sanford
Steever. Routledge publications. 40-74.

Cenoz, J. and Valencia, J. F. (1994). "Additive trilingualism: evidence from the Basque
country," Applied Psycholinguistics, 15, 195-207

Cenoz, J. (2001). "The effect of linguistic distance, L2 status and age on cross-linguistic
influence in third language acquisition," In Cross-Linguistic Influence in Third
LanguageAcquisition: Psycholinguistc Perspectives, edited by Cenoz, Hufeison & Jessner.
8-20.

Cenoz, J. (2003). "The additive effect of bilingualism on third language acquisition: A review,"
The International Journal of Bilingualism, 7 (1), pp. 71-87.

Costa, A. (2004). "Speech production in Bilinguals," In The Handbook ofBilingualism, edited by
T. K. Bhatia and W. C. Ritchie, Oxford: Blackwell publishing. 201-223.

Chatterji, S. K. (1921). Bengali Phonetics. Bulletin of the School of Oriental Studies, University
of London, Vol. 2 (1), 1-25. Cambridge University Press.

DeBot, K. (1992). "A Bilingual Production Model: Levelt's "Speaking" Model Approach,"
Applied Linguistics, 13(1): 1-24.

Dalby, J. and Kewley-Port, Diane. (1999). "Explicit pronunciation training using automatic
speech recognition technology," CALICO journal, 16 (3) 425-444.

Enomoto, K. (1994). "L2 perceptual acquisition: The effect of multilingual linguistic experience
on the perception of a "less novel" contrast," Edinburgh Working Papers in Applied
Linguistics, 5, 15-29.

Ferguson, C.A. and Chowdhary, M. (1960). "The Phonemes of Bengali", Language, 36, (1), 22-
59.

Flege, J. (1998). "The phonetic study of bilingualism," Ilha do Desterro, 35, 19-28.

Flege, J., MacKay, I., and Piske, T. (2002). "Assessing bilingual dominance," Applied
Psycholinguistics, 23, 567-598.

Flege, J.E., Schirru, C. and MacKay, I.R.A. (2003). "Interaction between the native and second
language phonetic subsystems," Speech Communication 40, 467 -91.









Flege, J.E. (1995). "Second language speech learning: Theory, findings and problems," Speech
Perception and Linguistic Experience: Issues in Cross-Language research, edited by W.
Strange 233-277.

Francis, A. L. and Nusbaum, H. C. (2002). "Selective attention and the acquisition of new
phonetic categories," Journal of experimental psychology: human perception and
performance. 28 (2) 349-366.

Gibson, E. J. (1991). An Odyssey in Learning and Perception. Cambridge, MA: Bradford Books
MIT Press.

Gibson, J. J. (1966). The Senses Considered as Perceptual Systems. Boston, MA: Houghton-
Mifflin.

Gibson, J. J. (1979). The Ecological Approach to Visual Perception. Boston, MA: Houghton-
Mifflin.

Gonzalez-Ardeo, J. (2000). "Engineering students and ESP in the Basque Country: SLA versus
TLA," In J. Cenoz, B. Hufeisen and U. Jessner (eds) Looking beyond second language
acquisition: Studies in tri- and multilingualism, pp 75-95, Stauffenburg.

Grosjean, F. (1998). "Studying Bilinguals: methodological and conceptual issues," Bilingualism:
language and cognition, 1 (2), 131-149.

Grosjean, F., Li, P., Munte, T.F. and Rodriguez-Fornells, A. (2003). "Imaging Bilinguals: when
the neuroscience meet the language sciences," Bilingualism: Language and Cognition 6
(2) 159-165.

Grosjean, F. (1992). "Another view of bilingualism," In Cognitive processing in bilinguals, edited
by R.J. Harris, Elsevier Science publishers 51-62.

Haldar, G. (1986). A Comparative Grammar of East Bengali Dialects, Calcutta: Puthipatra.

Halle, P. A., Best, C. T. and Levitt, A. (1999). "Phonetic vs. phonological influences on French
listeners: perception of American English approximants," Journal of Phonetics, 27, 281-
306.

Hammarberg, B. (2001). "Roles of LI and L2 in L3 Production and Acquisition," In Cross-
Linguistic Influence in Third Language Acquisition: Psycholinguistc Perspectives, edited
by Cenoz, Hufeison & Jessner, 21-41.

Herdina, P. and Jessner, U. (2002). A dynamic model ofMultilingualism. Clevedon: Multilingual
Matters.

Hoffmann, C. (2001). "Towards a description of trilingual competence," International journal of
bilingualism. 5 (1) 1-17.









Hammond, R.M. (2001). The Sounds of Spanish: Analysis and Application. Cascadilla Press.
2001.

Harnsberger, J.D. (1998) "The perception of non-native nasal contrasts: A cross-linguistic
perspective". PhD Dissertation, University of Michigan.

Iverson, P., Hazan, V. and Bannister, K. (2005). "Phonetic training with acoustic cue
manipulations: A comparison of methods for teaching English /r/-/1/ to Japanese adults," J.
Acoust. Soc. Am. 118, (5) 3267-78.

Jamieson, D., and Morosan, D. (1986). "Training non-native speech contrasts in adults:
Acquisition of the English /d/-/6/ contrast by francophones," Percept. Psychophys. 40, 205-
215.

Jamieson, D.G. & Morosan, D.E. (1989). Training new, non-native speech contrasts: a
comparison of the prototype and perceptual fading techniques. Canadian Journal of
Psychology, 43 (1), 88-96.

Jessner, U. (2006). Linguistic Awareness in Multilinguals: English as a Third Language.
Edinburgh University Press.

Jusczyk P. (1989). "Developing phonological categories from the speech signal," paper
presented at The International Conference on Phonological Development," Stanford
University.

Kalackel, T. (1985). A contrastive Analysis of the Phonological Systems of English and
Malayalam, Ph.D. Thesis, Northern Illinois University.

Klein, Elaine C. (1995). "Second versus third language acquisition: is there a difference?,"
Language Learning. 45 (3) 419-468.

Kostic, D. and Das, R. (1972). A .h\/,, t Outline of Bengali Phonetics, Calcutta: Statistical
Publishing Society.

Kroll, J. & Sunderman, G. (2003). "Cognitive processes in second language learners and
bilinguals: the development of lexical and conceptual representations," In The Handbook
of Second language Acquisition, edited by Doughty & Long, Oxford: Blackwell
Publishing, 104-129.

Krishnamurti, Bhadriraju. (2003). The Dravidian Language, Cambridge University Press 2003.

Kirsner K., Smith, M.C., Lockhart, R. S., King, M. L. and Jain M. (1984). "The bilingual
lexicon: language-specific units in an integrated network," Journal of verbal learning and
verbal behavior, 23, 519-539.

Krashen, S. (1981). Second Language Acquisition and Second Language Learning. Oxford:
Pergamon press.









Kroll, J. F., and Tokowicz, N. (2005). "Models of bilingual representation and processing:
looking back and to the future," In Handbook of bilingualism: psycholinguistic
approaches, edited by J. F. Kroll and A. M. B. Degroot, New York: Oxford University
Press 531-553.

Kruschke, J. (1992). "ALCOVE: An exemplar-based connectionist model of category learning,"
Psychol. Rev. 90 22-44.

Kuhl, P. K. (1991). "Human adults and human infants show a 'perceptual magnet effect' for the
prototypes of speech categories, monkeys do not," Percept. Psychophys. 50 93-107.

Kuhl, P. K., Williams, K. A., Lacerda, F., Stevens, K. N., and Lindblom, B. (1992). "Linguistic
experience alters phonetic perception in infants by 6 months of age," Science 255, 606-
608.

Kuhl, P.K. and Iverson, P. (1995). "Linguistic experience and the 'perceptual magnet effect'," In
Speech perception and linguistic experience: Issues in cross-language research, edited by
W. Strange, 121-154.

Kumari, B. S. (1972). Malayalam Phonetic Reader. Mysore: Central Institute of Indian
Languages.

Lively, S.E., Logan, J.S. and Pisoni, D.B. (1993). "Training Japanese listeners to identify English
/r/ and /1/. II: The role of phonetic environment and talker variability in learning new
perceptual categories," Journal of acoustical society of America, 94 (3) 1242-55.

Lively, S.E and Pisoni, D.B. (1994). "Training Japanese listeners to identify English /r/ and /1/.
III: Long-term retention of new phonetic categories," Journal of acoustical society of
America, 96 (4) 2076-2087.

Logan, J.S., Lively, S.E. & Pisoni, D.B. 1991. Training Japanese listeners to identify English /r/
and /1/: A first report. Journal of Acoustical society of America, 89 (2), pp. 874-886.

Ladefoged, P. and Maddieson, I. (1996). The Sounds of World's languages. Blackwell
Publishers.

Ladefoged, P. and Bhaskararao, Peri. (1983). "Non-quantal aspects of consonant production: A
study of retroflex consonants," University of California Working Papers in Phonetics, 57
11-25.

Ladefoged, P. (2005). Vowels and Consonants: an introduction to the sounds of languages.
Blackwell Publishing.

Ladefoged, P. (1999). English," In Handbook of the International Phonetic Association.
Cambridge Univeristy Press.

Lindau, Mona. (1985). "The Story of/r/," in Phonetic Linguistics: Essays in honor of Peter
Ladefoged, edited by V. A. Fromkin, Academic Press, Inc.









Major, R. and Kim, E. (1996). "The Similarity Differential Rate Hypothesis," Language
Learning 46 (3) 465-496.

Malakoff, M. (1992). "Translation Ability: A Natural Bilingual and Metalinguistic Skill," in
Cognitive Processing in Bilinguals, edited by R. Harris, Amsterdam: North Holland, 515-
30.

McAllister, R., Flege, J.E. and Piske,T. (2002). "The influence of LI on the Swedish quantity by
native speakers of Spanish, English and Estonian," Journal of Phonetics, 30, 229-258.

McDonough, J. and Johnson, K. (1997). "Tamil Liquids: an investigation into the basis of the
contrast among five liquids in a dialect of Tamil," Journal of International Phonetic
Association, 27, 1-2.

Mohanan, K. P. and Mohanan, T. (1984). "Lexical Phonology of the consonant system in
Malayalam," Linguistic Inquiry, 15 (4) 575-602.

Mufioz, C. (2000). "Bilingualism and trilingualism in school students in Catalonia," In J. Cenoz
and U. Jessner (eds.) English in Europe: The acquisition of a third language, pp. 157-178,
Clevedon: Multilingual Matters.

Narayanan, S., Kaun, A., Byrd, D., Ladefoged, P. and Alwan, A. (1996). "Liquids in Tamil,"
ICSLP, Philadelphia, PA.

Narayanan, S., Byrd, D. and Kaun, A. (1999). "Geometry, kinematics and acoustics of Tamil
liquid consonants," Journal of Acoustical Society of America, 106(4).

Neri, A., Cucchiarini, C., Strik, H. and Boves, L. (2002). "The pedagogy-technology interface in
computer assisted pronunciation training," Computer Assisted Language Learning, 15 (5)
441-467.

Nosofsky, R. M. (1986). "Attention, similarity, and the identification categorization
relationship," J. Exp. Psychol. Gen., 115 39-57.

Nosofsky,R. M. (1987). "Attention and learning processes in the identification and
categorization of integral stimuli," J. Exp. Psychol. Learn. Mem. Cogn. 15 700-708.

Olive, J.P., Greenwood, A. and Coleman, J. (1993). Acoustics ofAmerican English Speech: a
dynamic approach. New York: Springer-Verlag.

Peal, E. and Lambert, W. (1962). "The relation of bilingualism to intelligence," Psychological
Monographs, 76, 1-23.

Polka, L. (1991). "Cross-Language speech perception in adults: Phonemic, Phonetic and acoustic
contributions," Journal of Acoustical Society of America. 89 (6).

Polka, L. (1992) "Characterizing the influence of native language experience on adult speech
perception", Perception & Psychophysics. 52 (1) 37-52.









Pruitt, J.S, Jenkins, J.J. and Strange, W. (2006). "Training the perception of Hindi dental and
retroflex stops by native speakers of English and Japanese," Journal of Acoustical society
of America. 119 (3).

Raja, K. K. (1960). "The nasal phonemes of Malayalam," Indian Linguistics, 21 90-96.

Ray, P. S., Hai, M. A., and Ray, L. (1966). Bengali Language Handbook, Washington D. C.,
Center for Applied Linguistics.

Roach, P. (1983). English phonetics and phonology. Cambridge: Cambridge University Press.

Sanz, C. (2000). "Bilingual education enhances third language acquisition: evidence from
Catalonia," Applied Psycholinguistics. 21, 23-44.

Silver, J. (2007). "The organization of the bilingual perceptual space: English/Spanish bilingual
perception of Malayalam nasal consonants," Masters thesis, University of Florida.

Smiljanic, R. and Bradlow, A. (2007). "Clear speech intelligibility: Listener and talker effects,"
ICPhS XVI, Saarbrucken 6-10 August 2007, 661-664.

Smiljanic, R. and Bradlow, A. (2005). "Production and perception of clear speech in Croatian
and English," J.Acoust. Soc. of Am., 118(3), 1677-1688.

Sreedhar, M. V. (1972). "Phonology of the Cochin dialect of Malayalam," International Journal
of Dravidian Linguistics 1.2 100-125.

Strange, W. and Dittmann, S. (1984). "Effects of discrimination training on the perception of/r-l/
by Japanese adults learning English," Percept Psychophys. 36 (2) 131-145.

Strange, W., Polka, L. and Aguilar, M.C. (1989). "Effects of auditory and phonetic training on
Americans' discrimination of Hindi retroflex dental contrasts," Journal of Acoustical
Society of America, 86 (Sl) S101.

Strange, W. (1995). Speech Perception and Linguistic Experience: issues in cross-language
research. Baltimore: York Press.

Subramoniam, V. I. (1973). Note on the dental and alveolar nasal reflex as in Malayalam,"
Dravidian Linguistic Conference III, Dharwad, India.

Sundara, M., Polka, L. and Genesee, F. (2006). "Language-experience facilitate discrimination
of/d-6/ in monolingual and bilingual acquisition of English," Cognition, 100 369-388.

Sundara, M. and Polka, L. (2007). "Discrimination of coronal stops by bilingual adults: The
timing and nature of language interaction," Cognition 106 234-258.

Tajima, K., Kato, H., Rothwell, A., Akahane-Yamada, R., Munhall, K. G. (2008). Training
English listeners to perceive phonemic length contrasts in Japanese," J. Acoust. Soc. Am.
123 (1) 397-413.









Tees, R. C. and Werker, J. F. (1984). "Perceptual flexibility: maintenance or recovery f the
ability to discriminate non-native speech sounds," Canadian Journal of Psychology 38 579-
590.

Tunga, S. S. (1995). Bengali and Other Related dialects of Sn,,ih Assam, New Delhi: Mittal
Publications.

Uchanski, R. M. (2000). "Clear speech," In The Handbook of Speech Perception, edited by D.B.
Pisoni and R.E. Remez, Blackwell Publishing 207-235.

Werker, J. F., Gilbert, J. H. V., Humphrey, K., and Tees, R. C. (1981). "Developmental aspects
of cross-language speech perception," Child Development 52 349-353.

Werker, J.F. and Tees, R.C. (1984). "Phonemic and phonetic factors in adult cross-language
speech perception," Journal of Acoustical Society of America 75 (1).









BIOGRAPHICAL SKETCH

Divya Verma Gogoi received her Ph.D. in linguistics from the University of Florida in

2010. During her time at UF, she worked as a teaching assistant for the Department of

Linguistics and worked as a research assistant to Dr. Caroline Wiltshire. Her doctoral degree was

supported by a grant from Language Learning: A Journal of Research in Language Studies.

Before enrolling as a graduate student at UF, she received her B.A and then M.A in English from

the Panjab University, Chandigarh, India in 1999 and 2001 respectively. Thereafter, she

completed her M.Phil in Linguistics at the Central Institute of English and Foreign Languages,

India (now The English and Foreign Languages University) in 2005 with specialization in

Phonetics and Phonology. Her research interests are Phonetics and second/third language

acquisition; more specifically, acoustic phonetics and cross-language speech perception and

learning.





PAGE 1

1 ACQUISITION OF NOVEL PERCEPTUAL CATEGORIES IN A THIRD LANGUAGE: THE ROLE OF METALINGUISTIC AWARENESS AND FEATURE GENERALIZATION By DIVYA VERMA GOGOI A DISSERTATION PRESENTED TO THE GRADUATE SCHOOL OF THE UNIVERSITY OF FLO RIDA IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF DOCTOR OF PHILOSOPHY UNIVERSITY OF FLORIDA 2010

PAGE 2

2 2010 Divya Verma Gogoi

PAGE 3

3 To my parents m y husband, Ankur and our daughter Kaavya

PAGE 4

4 ACKNOWLE DGMENTS Working on this dissertation has been a challenging and enriching experience for me. me to undertake this research. I would first like to thank my academi c advisor and committee chair, Dr. Caroline Wiltshire and my Co Chair, Dr. James Harnsberger, for their guidance constant encouragement and willingness to be available for discussions, t heir incise comments and patient feedback, even if I went off track s everal times have proved invaluable to this work. I really appreciate Dr. I thank my dissertation committee members Dr. Wayland, Dr. Lord and Dr. Shrivastav for lending their invaluable gui dance as phonetic experts and language specialists towards my di ssertation work. To Dr. Wayland for making international students feel welcomed during our initial stay at UF and her constant willingness to give feedback whenever sought. I would like to tak e this opportunity to thank the Language Learning dissertation Grant Committee for approving the partial funding for this project. I express my gratitude to all the subjects who participated in this research Without their co n able to complete the project in time. To my parents month which never came. They have always supported me to do what I feel is right and then be ready to face the consequences, good or bad. I thank Bis ha, for her immense support in every way during the data collection process in India. I thank all the wonderful people with whom I had a chance to be friends with at UF Especially, Monica, Ashwini, Srikant and my fellow graduate students: Hanye, Weihua Zh u, Zoe Ziliak, Sharon Barkley and Quinn Hansen who have been a big support to me and most importantly I would like to thank my husband Ankur for being a constant source of energy that fuelled me to take on a lot of things unbridled.

PAGE 5

5 TABLE OF CONTENTS page ACKNOWLEDGMENTS ................................ ................................ ................................ ............... 4 LIS T OF TABLES ................................ ................................ ................................ ........................... 8 LIST OF FIGURES ................................ ................................ ................................ ......................... 9 ABSTRACT ................................ ................................ ................................ ................................ ... 11 CHAPTER 1 INTRODUCTION ................................ ................................ ................................ .................. 13 Overview ................................ ................................ ................................ ................................ 13 Goals of the study ................................ ................................ ................................ ................... 14 Literature Review ................................ ................................ ................................ ................... 15 Third Language Acquisition ................................ ................................ ............................ 15 Factors Influencing Acquisition of L2/L3 Speech Sounds ................................ .............. 17 Propos ed Factors ................................ ................................ ................................ ............. 19 Metalinguistic awareness ................................ ................................ ......................... 19 Feature generalization ................................ ................................ .............................. 21 Perce ptual assimilation ................................ ................................ ............................. 23 Cross Language Speech Perception Models ................................ ................................ ........... 25 Speech Learning Model (SLM) ................................ ................................ ....................... 25 Perceptual Assimilation Model (PAM) ................................ ................................ ........... 27 The Present Study ................................ ................................ ................................ ................... 31 Organization ................................ ................................ ................................ ........................... 34 2 METHODOLOGY ................................ ................................ ................................ ................. 35 Introduction ................................ ................................ ................................ ............................. 35 Description of languages ................................ ................................ ................................ ........ 35 Malayalam ................................ ................................ ................................ ....................... 36 Nasals ................................ ................................ ................................ ....................... 37 Fricatives ................................ ................................ ................................ .................. 39 Laterals ................................ ................................ ................................ ..................... 37 Rhotics ................................ ................................ ................................ ...................... 40 Bengali ................................ ................................ ................................ ............................. 41 Spanish ................................ ................................ ................................ ............................ 43 American English ................................ ................................ ................................ ............ 45 Stimulus Materials ................................ ................................ ................................ .................. 46 Stimulus Material Evaluation ................................ ................................ .......................... 49 Stimuli Evaluation: Preliminary Experiment ................................ ................................ .. 50 Stimulus Preparation ................................ ................................ ................................ ............... 51 Subjects ................................ ................................ ................................ ................................ ... 54

PAGE 6

6 Bengali English Bilinguals ................................ ................................ .............................. 56 Spanish English Bilinguals ................................ ................................ ............................. 57 American English Monolinguals ................................ ................................ ..................... 58 Experimental Design ................................ ................................ ................................ .............. 59 Perception Training Method ................................ ................................ ............................ 59 Procedure ................................ ................................ ................................ .................. 61 Familiarization ................................ ................................ ................................ ......... 62 Pretest ................................ ................................ ................................ ....................... 62 Training ................................ ................................ ................................ .................... 65 Posttest ................................ ................................ ................................ ..................... 66 Task Order Presentation ................................ ................................ ........................... 66 Data Analysis ................................ ................................ ................................ ................... 67 Summary ................................ ................................ ................................ ................................ 69 3 RESULTS ................................ ................................ ................................ ............................... 70 Introduction ................................ ................................ ................................ ............................. 70 Results ................................ ................................ ................................ ................................ ..... 72 Main Effects and Interaction of Independent Variables ................................ .................. 72 Examining Evidence for Individual Factors ................................ ................................ .... 76 Multili ngual benefit ................................ ................................ ................................ .. 76 Feature generalization ................................ ................................ .............................. 81 Re Analysis ................................ ................................ ................................ ............................. 84 Re analysis: Mu ltilingual Factor ................................ ................................ ..................... 85 An Alternative Re analysis: Only Place Contrasts ................................ .......................... 88 Perceptual Assimilation Results ................................ ................................ ............................. 92 Changes in Assimilation Types after Training ................................ ................................ 93 Changes in Assimilation Types: Multilingual Benefit Factor ................................ ......... 95 Summary ................................ ................................ ................................ ................................ 99 4 DISCUSSION AND CONCLUSION ................................ ................................ .................. 101 Introduction ................................ ................................ ................................ ........................... 101 Ov erview of the Results ................................ ................................ ................................ ........ 101 Alternate Explanations ................................ ................................ ................................ .......... 104 Test of Generalization Results ................................ ................................ ....................... 105 Necessity for Reanalysis ................................ ................................ ................................ 107 Caveats in Perceptual Assimilation Analysis ................................ ................................ 111 Results in Relation To Previous Studies ................................ ................................ ............... 119 Conclusion ................................ ................................ ................................ ............................ 125 APPENDIX A PRELIMINARY SCREENING ................................ ................................ ............................ 127 B SUBJECT INFORMATI ON ................................ ................................ ................................ 130 C EXPERIMENT INSTRUCTION SHEET ................................ ................................ ............ 133

PAGE 7

7 D PERCEPTUAL ASSIMILATION RESPONSE SHEET ................................ ..................... 136 E RESULTS IN TABULAR FORMAT ................................ ................................ .................. 137 LIST OF REFERENCES ................................ ................................ ................................ ............. 139 BIOGRAPHICAL SKETCH ................................ ................................ ................................ ....... 147

PAGE 8

8 LIST OF TABLES Table page 2 1 Phonemic consonant inventory of Malayalam ................................ ................................ ... 37 2 2 Phonemic consonant inventory of Bengali ................................ ................................ ........ 42 2 3 Phonemic consonant inventory of Spanish ................................ ................................ ........ 44 2 4 Phonemic consonant inventory of American English ................................ ........................ 45 2 5 Talker information; all talkers are native speakers of Malayalam and born in the state of Kerala, India; ................................ ................................ ................................ ................. 48 2 6 Result of the preliminary AX discrimination task across thre e Malayalam speakers. Mean D prime values averaged across contrasts as well as across vowels. ..................... 51

PAGE 9

9 LIST OF FIGURES Figure page 2 1 A map of the state of Kerala with cities of origin for the Malayalam speakers. A smaller map of India on the left hand bottom corner shows the location of the state of Kerala. ................................ ................................ ................................ ................................ 49 2 2 An outline map of Mexico, Central America and South America showing the regional distribution of the subjects. ................................ ................................ .................. 58 2 3 Different trials possible in an AX discrimination task using two tokens of one speech sounds. ................................ ................................ ................................ ............................... 64 3 1 Mean percent correct identification scores averaged over all four contrasts at the pretest and the posttest level. ................................ ................................ ............................. 73 3 2 iscrimination test over all four contrasts at the pretest and ................................ ............................ 74 3 3 Mean percent correct identification of contrasts spoken by trained on talker (posttest) and new talker (generalization test). ................................ ................................ .................. 76 3 4 Individual comparison of the monolingual group with BE bilingual group (top chart) and SE bilingual group (bottom chart). ................................ ................................ .............. 77 3 5 Mean percent correct response in identifying the lateral non native contrast [ ] in the identification test at the pretest and posttest level. ................................ ....................... 79 3 6 native contrast [ ] at the pretest and posttest levels by the subjects in the three language groups. ................................ ...... 81 3 7 Mean percent identific ation scores for the nasal non native contrast [ ] by the subject groups at the pretest and the posttest level. ................................ ........................... 82 3 8 Mean percent identification scores for the fricative non native contrast [ ] by the three language groups at the pretest and the posttest level. ................................ ............... 83 3 9 Mean percent scores of identification test across all four contrasts by the bilingual groups (BE and SE) and the m onolingual group (AE) at the pretest and the posttest levels. ................................ ................................ ................................ ................................ 86 3 10 groups (BE and SE) and the monolingual g roup (AE) at the pretest and the posttest levels ................................ ................................ ................................ ................................ .. 87

PAGE 10

10 3 11 Mean percent correct response for identification of the three place contrast (laterals, nasals, fricatives) by bilingual (BE and SE) and m onolingual (AE) groups at the pretest and posttest levels. ................................ ................................ ................................ .. 90 3 12 (Laterals, nasals, fricatives) by the language gro ups at the pretest and posttest levels. .... 91 3 13 Mean percentage of assimilation elicited from all three language groups, pretest vs posttest. ................................ ................................ ................................ .............................. 94 3 14 Percentage of difference scores (pretest to posttest) of assimilation types for the individual contrasts elicited from all three language groups. ................................ ............ 95 3 15 Percentages of assimilati on types elicited from individual language groups at the pretest and posttest levels are provided in three different charts. ................................ ...... 97 3 16 Percentage of difference scores (posttest pretest) of assimila tion types elicited from BE, SE bilingual groups and AE monolingual group over all four contrasts. ................... 98 4 1 Hypothetical representation of the learning continuum displaying the ambiguity of ranking the UU, UC and CG assimilation types in order to assess the direction of learning. ................................ ................................ ................................ ........................... 118

PAGE 11

11 Abstract of Dissertation Presented to the Graduate School of the University of Florida in Partial Fulfillment of the R equirements for the Degree of Doctor of Philosophy ACQUISITION OF NOVEL PERCEPTUAL CATEGORIES IN A THIRD LANGUAGE: THE ROLE OF METALINGUISTIC AWARENESS AND FEATURE GENERALIZATION By Divya Verma Gogoi August 20 10 Chair: Caroline R. Wiltshire Co chair: Ja mes D. Harnsberger Major: Linguistics M any factors that contribute towards the perception of second language (L2) sounds and the subsequent establishment of L2 phonetic categories ha ve been the focus of past research However, learning a third language, u nlike SLA, may be influenced by additional factors attributed to the presence of two language systems in a bi lingual instead of one language system in a monolingual. Previous third language acquisition research mainly lexical processing or syntax acquisit ion, have suggested that being a bilingual provides a positive influence to learners of a third language in attaining general proficiency. Whether b ilingualism or a having multilingual benefit as a factor influences the acquisition of novel non native spee ch contrasts by adult bilinguals, remains to be explored. Another factor that may facilitate in the phonetic acquisition of a third language is the role that phonetic features may play in the development of new phonetic categories. F eatures utilized in nat ive contrasts may generalize in the learning of novel non native contrasts, even if they play a limited or no role in the initial perception of these non native contrasts Such factors influencing third language acquisition within specific areas of profici ency like phonological acquisition have not been previously investigated. Th is study involves a perceptual training experiment examining the influence of factor s of m ultilingual benefit and/or feature generalization The training incorporates the acquisiti on of

PAGE 12

12 four speech contrasts (dental/alveolar retroflex distinction) of a third language by bilingual (Bengali English and Spanish English speakers) and monolingual (American English speakers) groups The perceptual data is gathered through measures like consonant identification, AX discrimination and perceptual assimilation. The perceptual assimilation data is analyzed using the PAM model to determine the extent to which the initial assimilation patterns change towards direction of learning following the training. The results across the testing measures show subtle effects of the multilingual benefit factor playing a role in facilitating the acquisition of the non native speech contrasts when comparing individual language groups This effect is observed as robust when both bilingual groups are combined to increase sample size during reanalysis in all tests, except for AX discrimination The factor of feature generalization does not show an influence on the acquisition of non native speech contrasts by Benga li English bilinguals in any of the testing measures. Therefore, the results are suggestive of only the multilingual benefit factor playing a role in enhancing the acquisition process of a third language by adult bilinguals The perceptual assimilation ana lysis across language groups show a change of assimilation types after training revealing a shift towards the direction of learning. The bilinguals show a greater shift from SC assimilation type to TC assimilation types than monolinguals. The study thus su ggests that multilingualism as a factor facilitates the acquisition of non native speech contrasts.

PAGE 13

13 CHAPTER 1 INTRODUCTION Overview The globalization of the world has led to many scenarios where people with different native languages come face to face wi th each other and the need to communicate arises The field of second language acquisition at the cognitive level has a long history of exploration of the issue of whether bilinguals have a single system of memory representation or separate systems for eac h language that they use. Previous studies on this issue have suggested presence of a single or shared system for more than one language at the lexicon level but acknowledge the probability of the presence of more language specific levels as well ( see Fleg e 1995 ; Kirsner et al. 1984 ; Grosjean 1992 ; Krashen, 1981 ; Kroll and Tokowicz 2005 ; De Bot 1992 ). SLA studies have looked into various factors that influence the process of acquiring a new language by adults who have only one language system However, the scenario is not that simple. Acquisition of more than two languages has become more a norm than exception. The refore, research in the field of third language acquisition or in broader perspective, multilingualism has made strides in the last few years Since the last decade research ers have turned their attention to investigating the factors that influence the acquisition of a third language (Cenoz, Hufeisen & Jessner 2001 ; Baker 2001 ; Bialystok 2001; Munoz 2000 ; Sanz 2000). While there has been g rowing interest in L2 acquisition and bilingualism, many studies have looked into the factors such as age of acquisition, amount of use of language, context, status of the L1 etc. and their effect on learning L2 either simultaneous or sequential. Fewer stu dies have looked into third language acquisition within specific area s of proficiency like phonological acquisition (Enomoto 1994 ; Werker 1986). Previous third language acquisition research have suggested that being bilingual provides

PAGE 14

14 a positive influenc e to learners of third language in attaining general proficiency (Cenoz & Valencia 1994 ; Ardeo 2000 ; Munoz 2000 ; Bi l d & Swain 1989). These studies were mostly done on children in classroom setting s Whether or not factors such as metalinguistic awarene ss and/or feature generalization enhance the acquisition of the phonetics of a third language for the bilinguals has not been previously investigated. In the present study, a n experiment is set up in this study to examine such factors after providing train ing with the new contrast in a laboratory setting. Goals of the study The goal of the present study concerns the exploration of certain factors that may positively influence the acquisition of a third language by adult bilinguals. Specifically, it examines the effects of bilingualism in terms of two factors; firstly, benefits of possessing two linguistic systems often termed as metalinguistic awareness and secondly, transfer of phonetic feature knowledge from L1/L2 to L3. The study investigates the effects of these two factors on the acquisition of a third language, Malayalam by two bilingual groups Bengali English and Spanish English speakers and one monolingual group of American English speakers. By presenting the subjects with novel speech contrasts (retr oflex sounds in manners that are lacking examine if any additive effects of being a bilingual are seen in the course of learning. The specific sound contrasts fro m Malayalam that are used as stimuli are lateral [l ], nasal [n ], fricative [ ] and rhotic [ ]. The experiment set up for this study includes various tests of identification, discrimination and perceptual assimilation to observe the effects of the factors in different levels of perception.

PAGE 15

15 The main hypothesis being tested in the present study is that the bilingual groups are expected to perform better than the monolingual group in perceiving and identifying the speech contrasts from the target la nguage. Such results would argue for positive influence of bilingualism on the third language acquisition in the specific area of phonetics. Within the bilingual groups the Bengali English learner group, who have phonemic experience with retroflex feature but only in manners other than what is presented in the experiment, is expected to perform better than the Spanish English learner group, who have no exposure to retroflex feature, phonemic or phonetic. By incorporating previous language experience as a f actor within the bilingual groups the results will provide information for an additive effect of feature generalization in enhancing the acquisition of speech contrasts from third language over the effect of bilingualism. The results will also be analyzed in terms of whether they fit the predictions of the Perceptual Assimilation Model (PAM) proposed by Best & Tyler 2007. T he study aims to examine the changes in assimilation patterns of nonnative speech contrasts by the listener groups at both the initial perceptual stage and after training stage by adopting the assimilations types described in PAM Literature Review Third Language Acquisition A vast corpus of literature exists concerning second language acquisition (SLA) in adulthood. Research on third lan guage acquisition as a separate entity is a relatively recent endeavor. Therefore, research in this field is not as extensive. Third language acquisition refers to the learning of a non native language by an individual who has already acquired two language s either simultaneously (before puberty) or sequentially. The scenario of third language acquisition although quite similar to SLA still presents differences since, unlike second language learners, third language learners have more experience as language l earners and have access to

PAGE 16

16 two linguistic systems when acquiring a third language ( see Cenoz 2003) Most of the p revious studies on third language acquisition have looked into the question of whether bilingualism facilitate the acquisition of third langua ge? (Cenoz, Hufeisen & Jessner 2001 ; Baker 2001 ; Bialystok 2001; Munoz 2000 ; Sanz 2000) Most of these studies have been conducted on school children in a classroom setting. The focus of these studies have been primarily to examine the level of general proficiency in the target language being learnt through variou s linguistic tasks like word awareness, grammar, speaking, reading and writing. Their results have suggested that being a bilingual provides positive influence on the acquisition of the third language ( Cenoz & Valencia 1994 ; Ardeo 2000 ; Munoz 2000 ; Bild & Swain 1989) In addition to the effect of bilingualism, t hird language acquisition also presents a diverse range of other factors that may influence the learning of the target language. Th ese factors are quite similar to t he typical factors that affect acquisition of the second language such as age of acquisition or the amount of use of the language ( see e.g. Flege 1998 ; Flege et al. 2002 ) or phonetic, phonemic or acoustic factors (Polka 1991) coupled with ot her influential factors that are traditionally explored in the field of bilingualism such as L1/L2 status, socioeconomic factors, degree of bilingualism etc (see Hammarberg 2001) More specific to the current study, the effect of b ilingualism also termed as metalinguistic awareness, along with feature generalization, here referred as the transfer of phonetic/phonemic feature knowledge from L1/L2 sound inventory to facilitate in the acquisition of L3 sounds, is examined in the area of phonological acquisition of third language. The following section provides a background review of certain factors that may influence acquisition of sounds in both L2 and L3 along with the d etailed review of the proposed factors metalinguistic awareness and feature generalization

PAGE 17

17 Factors I nfluencing A cquisition of L2/L3 S peech S ounds Through extensive research many factors that contribute towards the perception of second language (L2) sounds and subsequent establishment of L2 phonetic categories have be en examined Based on such factors, theories on c ross language speech perception by nave as well as experienced listeners have been developed ( e.g., p erceptual assimilation m odel, or PAM ( Best 1994, 2007; Flege 1995). The term nave listener here refers to a listener who has no prior exposure to the non native sounds of the target language and an experienced learner is acquainted with the non native sounds although he/she has not formed any phonetic categories corresponding to these sounds Phonetic, pho nemic and acoustic factors among others have been determined to influence the learning of L2 sounds (Polka 1991). Previous studies that have addressed such factors are briefly discussed below. T he factor of phonetic experience defined th specific phonetic segments 1991 : 2962), ha s been explored in cross language speech perception research. It has been hypothesized (Werker et al 1981 ; Tees and Werker 1984; Strange and Dittmann 1984 ; Polka 1992) that a non native contrast that contains a phonetic segment identical to a segment frequently experienced only in phonetic contexts by a subject in his language(s), is more easily differentiated by the learner as opposed to no n native contrasts which may not employ allophonic experience. The findings from increased difficulty in differentiating (after training) the [ ] place contrast compare d to the [ ] voicing contrast from Hindi. They credited allophonic experience as an explanation for (AE) facilitated the differentiation of the non native voic ing contrast. However, a limited number

PAGE 18

18 of subjects who had experience with the Hindi language in their early childhood were able to perceive the place contrast after training, suggesting early linguistic experience facilitates better perception of the pla ce contrast than the AE subjects with no early life experience with Hindi (Tees & Werker, 1984) e above given hypothesis but found no conclusive results due to presentation order effects. However, the allophonic experience hy pothesis was contested in subsequent studies (Best et al 1988 ; Pruitt et al 2006), suggesting that phonetic experience alone could not explain the variability found in perceptual performance on different non native contrasts and could even have a detri mental effect in the perception of non native contrasts. Phonemic and phonetic ex perience together ( presence of l inguistic experience ) are able to better explain the facilitation effects that are seen in the perception of non native contrasts (Best et al 1988 ; Sundara et al 2006). Another factor of acoustic salience that may influence the acquisition of L2/L3 sounds was proposed by Burnham (1986) to account for variable perception levels of non native contrasts. Burnham claimed that the distinction bet ween the easily discriminable and the poorly discriminable non native speech contrasts can be explained in terms of their inherent psychoacoustic basis. distinction between non native contrasts. According to him, a contrast was robust if the reliable acoustic cues in that speech contrast are too salient to be lost during memory demanding tasks. However, psychoacoustic salience individually as a factor can explain perceptual performance only when some independent acoustic cue measures can be determined. Further research needs to be conducted in this respect. The above discussed factors have been determined to substantially influence the perception of non native speech sounds of a target l anguage. However, without explicit definition, a target

PAGE 19

19 language can be assumed to be a second language (L2), a third language (L3) or a novel language (language unknown to the listener). The area of third language acquisition (or multilingualism) un til re cently was considered equivalent to second language acquisition (SLA) (as noted in Hammarberg 2001) However, learning a third language, unlike SLA, may be influenced by additional factors like effects of bilingualism (Hoffman 2001 ; Sanz 2000 ; Chenoz 2 001; Bialystok 1992, 2001) The section that follows provides detailed information on the factors explored and presents them in the context of the current study. P roposed Factors Metalinguistic a wareness Metalinguistic awareness allows the individual to s tep back from the comprehension of an utter ance; to attend to the structural features of the sounds in order to abstract the knowledge of its distinguishing cue s in a speech contrast a task which demands certain cognitive and linguistic skills (Malakoff 1992 : 518 ). This awareness helps a bilingual to acquire non native contrasts more easily than monolinguals. The assumptions in the formulation above are based on th e results of studies on lexical processing and word learning by bilingual and monolingual ch ildren showing that bilingual children are more adept at doing metalinguistic tasks and have cognitive benefits from knowing two or more languages (Peal and Lambert 1962; Munoz, 2000; Klein 1995; Sanz 2000 etc.). The study by Klein (1995) showed that dur ing lexic al acquisition, multilinguals learned the same number of lexical items as monolinguals but at a quicker rate of acquisition Klein based the explanation for this phenomenon on enhanced cognitive skills in multilinguals which helped them tease out the potential relevant data for resetting the particular parameter preposition stranding in this case, for the new language. Numerous studies looking at

PAGE 20

20 cognitive processing 1 at a linguistic level have explored the concept of metalinguistic awareness in b ilingual and multilingual children as well as adults (Bialystok 2004). The studies emphasize the fact that bilinguals with near native proficiency in both languages would show evidence of metalinguistic awareness (Bialystok 1992, 2001; see also review in Jessner 2006). In her study, Sanz (2000) suggests that metalinguistic awareness may be an enhancing factor in language processing. She promotes bi literacy as the cause of heightened metalinguistic awareness. m exposure to literacy in two languages, gives (Sanz 2000: 36). This study extend s the concept of metalinguistic ability from the field of lexical processing to the field of phonetics. This core concept of metalinguistic awareness has not yet been explored in the field of cross language speech perception and in the area of phonological acquisition of a third language It is yet to be seen if implicit knowledge of two languages, in the case of phonetic features in the sound systems, can result in heightened capacity to acquire non native speech sounds. However, the concept of metalinguistic awareness is a very broad usage for various metalinguistic s kills. Th is study specifically assesses metalinguistic awareness / metalinguistic ability from the perspective of the effects of bilingualism where having two or more languages well established within the language system of an individual benefit him for linguistic pro cessing over a monolingual individual. Therefore, rather than using the term metalinguistic awareness loosely, the more operational term of m ulti lingual benefit factor will be used in this study. According to previous studies in lexical processing, t he bil inguals may function at a faster rate and at a higher level for reorganizing information in order to deal with the high demand task of 1 For this study the term cognitive processing is operationally defined as the act of abstraction of relevant information from a speech utterance, required for a linguistic skill, as a result of higher level of concept formation.

PAGE 21

21 acquiring new non native speech contrasts when compared with monolinguals. The assumption that bilinguals develop meta li nguistic awareness which benefits them over monolinguals in acquiring a third language, can be examined through the experimental design set in which bilinguals are compared to monolinguals in the acquisition of speech sounds in the third language as is do ne in this study For instance, in the current study scenario, the retroflex feature which is non existent in the Spanish may still be acquired by t he m at a relatively faster rate of acquisition than English monolin guals. This may reflect that bilinguals employ the ir metalingual knowledge to pay attention to distinctions in auditory acoustic stimuli and to tease apart relevant acoustic cues in order to accurately distinguish between the newly acquired speech contrast s. The results of perceptual p erformance after training would be compared to results of the monolingual group. Therefore, in case of successful and faster acquisition, results would suggest the presence of meta linguistic awareness among bilinguals (Spanis h English) which facilitates the acquisition of a novel contrast without the limitation of feature generalization of an existing feature. Feature g eneralization F eatures, an important set of information contained in the established phonetic category, play a crucial role in the development of new phonetic categories (Nosofsky 1986,1987; Kruschke 1992; Lively et al. 1993; Francis and Nusbaum 2002). Features may refer to phonemic attributes of the sound system of a language, phonetic features or acoustic c ues attached to the specific segments of the system. According to Francis and Nusbaum (2002) f eatures can be stretched or shrunk according to relative weights given to them. For this study, features refer to the phonetic features within a phonemic segment of a language. The factor of f eature generalization or feature productivity as defined by Polka (1992 : 40 experience with a particular place distinction facilitates perception of the same phonemic place

PAGE 22

22 distinction in a different manner contex The factor of f eature generalization operationally defined for this study, refers to the concept where the listener being sensitive to individual features in his/her native phonetic categories is able to identify the same feature in a novel non native contrast and thus develop new phonetic categories. Few studies have looked at the feature generalization phenomenon in the field of cross language speech perception. Polka (1992) and Harnsberger (1998) both found results contrary to the expected results i n this study that feature generalization will reveal better perceptual performance. However, one factor that could explain their results is that these previous studies involved only the perceptual presentation of new non native contrast to nave listeners rather than conducting a perceptual training experiment and subsequently observing any effects of feature generalization during a perceptual testing phase. The p resent study is different from previous work s in that it takes the next logical step and exami nes the role of feature generalization in the perceptual performance data (posttest data) of TRAINED listeners as opposed to nave listeners ( see Polka 1992 ; Harnsberger 1998). Recently, a study by McAllister, Flege & Piske (2002) found that experience with the duration (phonemic lengthening) more than that of native Spanish speakers (both subject groups had lived in Sweden for at least 10 years) The difference between study and others was that the subjects were experienced learners of Swedish and used the target language often. This provides support for the hypothesis of feature generalization. To test for feature generalization in th e current study a group of Bengali English bilinguals (BE) undergo perceptual train ing to acquire the speech contrasts of retroflex nasal as opposed to alveolar nasal. The Bengali language has extensive use of the retroflex feature for stops in its phonemic inventory bu t lacks this feature in manners like nasal stop [ ], lateral approximant [ ] fricative [ ] and central

PAGE 23

23 approximant [ ] An o bservation of faster acquisition of th ese novel sound s by the BE bilingual group when compared with either the bilingual Spanish En glish group or the monolingual American English group, will indicate a successful generalization of the feature retroflex from stops to novel manners of articulation like nasal stops. The results of training on the identical stimuli from the group of Engli sh monolinguals would be observed as a control group. One criticism on the choice of subject group could be that the phenomenon of feature generalization could be evident in monolingual Bengali speakers as well. The current study deals with feature general ization as a factor of enhancing learnability of non native speech contrasts, regardless of whether the learner is a bilingual or monolingual. However, taking bilingual learners of L3 for this study provides us with the opportunity to assess both feature g eneralization and multilingual benefit within a single group of subjects. Moreover, if the results do not provide evidence for multilingual benefit (in the Spanish English (SE) group ) but show enhanced learnability in the Beng a li English group, it would su ggest that it is not the factor of multilingual benefit that is facilitating learnability in Bengali English but the fa ctor of feature generalization. On the contrary, if the multilingual benefit factor is observed in the SE group, then the feature general ization factor is assumed to have an additive effect that can be observed in BE group as significantly higher performance level than SE group. Perceptual a ssimilation p erceptual category inventory in the acquisition of non native contrasts from a third language. Looking at the perceptual assimilation patterns elicited from the learners at the post training stage will provide an opportunity to examine the formation of new perceptual categories if any, developed after a limited time of training. Moreover, perceptual assimilation analysis would be

PAGE 24

24 another dependent measur e to test for performance level of acquisition among the subject groups. This measure will show greater ecological validity as opposed to just perception tests conducted in previous studies that looked at the acquisition of L2/L3 through perceptual assimi lation. U nlike a perception test, training involves feedback on performance an essential aspect of learn ing Since a training experiment where an individual is perceptually trained on non native speech contrasts would simulate more closely the natural learning process, eliciting and examining perceptual assimilation (PA) types at the post training stage and comparing them with pre training PA types will prove to be more ecologically valid than simply eliciting the PA types in a perception experiment without any controlled training. Previously, s everal assimilation analytical models have been developed to acc ount for the influence of native perceptual categories on the perception, production, and acquisition of non native speech sounds, including the Speech Learning Model (SLM) and, of greater interest for this project, the Perceptual Assimilation Model (PAM) since it predicts the assimilation patterns not only at the initial exposure stage but also predicts learnability beyond the first stage Likewise, the current study also looks at the assimilation patterns at the post training stage. PAM concerns the discr iminability and learnability of non native contrasts based on the relationship between the non native and native category inventories These relationships have been encapsulated in several assimilation types: Two category (a non native contrast that is hig hly similar to a native contrast), Single category (a non native contrast that is consistently identified as a good exemplar of a single native speech sound), Category goodness (a non native contrast consistently of two speech sounds that are consistently identified with a single native sound but differ in goodness of fit), uncategorizable, either UC or UU (involving one/both non native speech sounds that are not consistently identified with a particular native category) and non

PAGE 25

25 assimilable (a contrast that is perceived in a nonlinguistic manner). These assimilation types have been discussed in detail in the next section. The present study tests these predictions of the PAM model in t wo ways: 1) Determining the extent to which initial patterns of assimilatio n change, if at all, following extensive training experience and 2 ) measuring the extent of assimilation pattern change in individual subject groups in order to assess any effects of the influencing factors explored in this study The following section dis cusses the above mentioned cross language speech perception models that have been developed predicting perceptual assimilation patterns of nave listeners as well as subsequent changes in assimilation patterns in the learning of a second language by experi enced learners. Cross Language Speech Perception Models Since the advent of speech perception studies on non native contrasts during the 1980 90s, various speech perception and acquisition models that address bilingualism have been proposed. This section o f the chapter describes the pertinent models and examines the core claims of each model. Speech Learning Model (SLM) The speech learning model focuses on the L2 phonological acquisition in production and perception (Flege 1995). It is claimed in this model that many second language production errors have an underlying perceptual basis. The model considers the factor of perception as the underlying principle behind the acquisition of L2 speech sounds and effective learning. The model propagates the notion th at if learners of second language sounds are not able to accurately perceive particular L2 sounds, the consequent production of those L2 sounds will be inaccurate. One of the preliminary assumptions of the model is that the production and perception of pho

PAGE 26

26 to accommodate L2 sounds, when encountered during the life span. Another postulate of this phonetic categories (long 1995: 239), bringing forth the relevance of the phonetic /phonemic contrasts between the L1 and L2 sounds, in the process of acquiring the second la nguage. Hypotheses of this model reflect the need to acknowledge the phonetic differences between the L1 and L2 sounds that may be phonetically closer or more distant, which may influence the perception of L2 sounds by the learners. The factor of mental re presentation or categorization of L1 and L2 sounds has been directly addressed in this model. According to SLM, if the bilinguals learn to distinguish some phonetic differences between the L2 sound and the closest L1 sound to it, a new phonetic category is likely to be established for that L2 sound. Nevertheless, the category formation for an other words, for the bilingual the L1 and L2 sounds can be perceptually li nked and therefore, one phonetic category is used for both sounds, which may eventually resemble each other in production. perception of these specific L1 and L2 soun ds by the listener as being quite similar (non perceivable phonetic distance by listener) (terms used interchangeably) sounds are easier because the listener easily distingu ishes the phonetic differences. SLM also predicts that if a new phonetic category for an L2 sound is in fact established, the phonetic system would try to maintain contrast between the existing L1 phonetic category and the newly formed L2 category since both L1 and L2 categories are assumed to com ( Flege 1995: 239)). Thus, the new L2 category established may

PAGE 27

27 not be identical to the native speaker category that is, the target sound Therefore, the predictions for mental representations of the L1 as well as L2 sounds are well documented in this model. The model differs from an older contrastive analysis approach as it proposes a less abstract phonetic level of analysis than phonemic level of analysis. Contrastive analysis of two languages involves the examination of phonemic d ifferences between the inventories of those languages. According to one SLM hypothesis, learners perceptually relate positional allophones in the L2 to the closest positionally defined allophone in the L1. On the other hand, L1 for L2 substitution describe d in contrastive analysis is seen as either failure to discern the phonetic differences between the L1 and L2 sounds or in ability to articulat e the correctly perceived sound. According to SLM, d uring the course of second language acquisition, the phonetic category that has already been established for an L1 sound during childhood evolves gradually if it is perceptually linked to an L2 sound. Such developments are influenced by age of learning and perceived phonetic distance between an L1 sound and the close st L2 sound. The model claims that the greater the perceived cross language phonetic distance, the more probability for the development of a separate phonetic category for the L2 sound. Also, age of learning (AOL) determines how much perceived phonetic dis tance is required to stimulate the process of separate category formation for the L2 sound (early AOL + smaller perceived phonetic distance needed = category formation). In summary, the model claims that the most L2 production errors observed have an under lying perceptual problem Perceptual Assimilation Model (PAM) The P erceptual A ssimilation M odel (Best 1995, 2007) lays its foundation on the phonetic/phonological nature of the L2 speech sounds versus L1 sounds, similar ly to the other speech perception mo dels. The preliminary assumption of the model is that the nature of the

PAGE 28

28 discrepancies and similarities that a non native listener picks up in the auditory signal may be ical D irect R ea list T heory (e.g., Gibson, 1991; Gibson 1966, 1979) simple gestural events in the speech signal (Best 1995). Th e model extends the direct realist theory in the way that the adult listener perceives information about gestural similarities in the L2 speech with the native phonemes, when receiving the non native signal. According to the model, there is no two extremes phenomenon (full assimilation or none) in perceiving non native speech. Also, all non native contrasts do not pose same level of difficulty to the listener. The perceptual difficulty should be predictable from the differences in the gestural similarity/di screpancy between non native and native sound distinctions. There are five main representation of non native speech contrasts. The predictions of the discrimination tas k are also given alongside the category type speech contrasts. Firstly, non native speech contrasts may be treated as two category type (TC) In this case, the speech contrasts may be gesturally similar to two different native phoneme categories making the non native speech contrast easily discriminable by the L1 listener. An adult listener is expected to show near native like discrimination of this type of non native speech contrast. Secondly, non native speech contrasts may be treated as single category t ype (SC) if native speech contrast is equally good or equally poor and both sounds are assimilated to a single category. Thirdly, non native speech sounds may show differences in category goodness type (CG). When the non na tive contrast is assimilated to a single category but one member of the contrast

PAGE 29

29 is seen as a better token of the native phoneme and the other as a marginal token of that category, this makes the discrimination easier than the SC type but not equivalent to TC type of assimilation. Again, because category goodness type of pairs may be perceived with varying degree of discrepancy between the two phones, the level of discrimination may vary. Finally, Uncategorizable type (UU) are formed where even though none of the phones from the non native pairs is assimilated with any native phoneme category, yet the phones are considered as speech sounds of a foreign language. For U ncategorized versus categorized type (UC) where only one phone of the non native pair is as similated to a native phoneme category, the adult listener is expected to show a very good discrimination performance for this type of non native speech contrast. Therefore, concluding from the above given possible outcomes of discrimination performance, t he pattern for an adult listener would be highest discrimination for Two category types moderate to good discrimination for both category goodness type, non assimilable type (for different reasons) as well as for the uncategorized versus categorized type and lowest for the single category type The uncategorizable type varies from good to poor depending upon the acoustic proximity of the two phones in the pair. Also, PAM assumes a strong L1 phonological system influence on the perception ability. The PAM model, in conclusion, emphasizes that the perception of non native segments is performed according to their similarities or discrepancies from the native segmental system, being in close proximity to the native phonological space as well as with each other (within the contrast) (Best 1995). Although the model provides intricate details of perceptual abilities of adults for non native speech contrasts with highly probable outcomes for the categorization of the non native phones by the listener, the model do es n o t really deal with the learning process of these non

PAGE 30

30 native phones by the L1 speaker, unlike SLM. The earlier PAM model provides the predictions of perception of non native speech contrasts by listeners who have n o t had any experience with L2. These d iscrimination outcomes provided in the model are based on the first exposure of the listener to these sounds. The learning process that takes place after the learner is exposed repeatedly to the L2 sounds and the reorganization of the L1 phonetic categorie s that takes place due to the influence of the L2 phones is not addressed in this model. These differences are addressed in the revised PAM model ( Best & Tyler 2007) The P erceptual A ssimilation M odel extends its predictions from the earlier model focusin g on the nave non native listener to a revised version of the model where it accounts for the experienced L2 learner apart from the initial encounter with L2 as a nave listener. Best refers to the revised version as PAM L2. Four possible cases of L2 mini mal contrasts are discussed starting with Only one L2 phonological category is perceived as equivalent (perceptually assimilated) to a given L1 phonological category (TC) T he discrimination of the speech contrasts would be excel lent as the L1 learner would always perceive the fully assimilated L2 phone separate from other less assimilated L2 phones. The L2 learning for this L2 phonological category will not take place as it has been entirely assimilated and recognized with an L1 phonological category. Both L2 phonological categories are perceived as equivalent to the same L1 phonological category, but one is perceived as being more deviant that the other (CG) The discrimination level for this type would range from moderate to goo d. During the learning process, the learner is likely to gradually form a new phonological category for the perceived poor/deviant phone of the L1 phonological category. The good exemplar L2 phone may remain assimilated to the L1 phonological category. Bot h L2 phonological categories are perceived as equivalent to the same L1 phonological category, but as equally good or poor instances of that category (SC) The listener will poorly discriminate the contrast since the two phones would be perceived as havin g very little phonetic difference and be considered as exemplars (either equally good or bad) of single L1 category. For the learner s to distinguish these two phones apart, relevant phonetic differences between the phones will have to be learnt, form ing a s eparate phonological category over time

PAGE 31

31 No L1 L2 phonological assimilation: There could be two situations with these types of minimal contrasts. The listener may not perceptually assimilate either of the phones with any L1 phonological category. If the tw o L2 phones are in existence with relatively larger listener. However, if the L2 phones are relatively close to each other in the L1 phonological space, the listen er would be not able to discriminate them easily. The PAM L2 model makes the assumption that the learner begins the acquisition of L2 as a nave but phonologically sophisticated L1 listener. With the initial introduction of the phones of L2 and new L2 phon es perceptually assimilated with L1 phonological/phonetic categories, the learning process for the L2 learner starts. The PAM L2 model makes predictions for the learning process of this, now experienced, L2 learner. The target phonological system, that the learner may achieve over a life span, which the model hypothesizes, is that of a common L1 and L2 system whereby phonetic categories from both L1 and L2 share a common phonological space All these models of cross language speech perception look at the se cond language learning by adults. These models can be extended to look at the factors facilitating third language learning. The current study proposes factors like metalingual benefit and feature generalization as some of those factors that may facilitate third and/or second language acquisition The Present Study This study is an attempt to provide a window to phenomena not explored well in the field of cross language speech perception and phonetics. Firstly, t he multilingual benefit concept has been attes ted only in psycholinguistics, lexical and cognitive processing or word learning fields. Moreover, un til recently multilingual benefit was examined only in bilingual children. The current study will test on in the area of cross language speech perception and third language learning Secondly the phenomenon of feature generalization has been evaluated in studies only dealing with nave listener perceptual performance. The main experiment of the study dea ls with training the subjects and testing their

PAGE 32

32 perceptual performance in order to seek evidence for feature generalization. Observing the perceptual performance of a t rain ed listener rather than a nave listener may be the missing factor in finding the ro le of feature generalization in facilitating the learnability of non native phonetic categories. Apart from the pioneering steps in exploring metalinguistic ability and feature generalization, the study adopts the assimilation types established by the revi sed PAM model and test s the predictions laid out by a ddressing the issue of the relation of initial perception modifying into different assimilation patterns during the course of training. The relevance of chosen stimulus set and the appropriate subject gr oups is discussed as follows. A stimulus set which is relatively difficult to discriminate based on acoustic differences is archetypal for this kind of training study so as to bring out subtle distinctions in the perceptual performance of the various subje ct groups. Dental /alveolar retroflex place distinction as a speech contrast is considered to have a reasonable level of difficulty in discrimination due to differences in spectral features as opposed to temporal features in speech contrasts (Pruitt et. al. 2006; Polka 1991 ; Strange and Dittmann 1984). Therefore, dental/alveolar retroflex place distinction was chosen as the place contrast for the non native speech contrast as the stimulus set. The speech contrasts were taken from Malayalam recorded as na tural stimuli from eight speakers Malayalam was chosen as the target language as it has a rich inventory of consonants employing the retroflex feature extensively in various manners. This allows a flexible choice of speech contrasts as per the requirement s of the study. An a ppropriate and diverse range of challenging non native speech contrasts were selected from Malayalam for the study. 1) [ ] ( alveolar versus retroflex lateral approximant) 2) [ ] (dental /alveolar versus retroflex nasal) 3) [ ] (p alat o alveolar versus retroflex fricative) 4) [ ] (alveolar tap versus retroflex approximant)

PAGE 33

33 These speech contrasts differ primarily in place of articulation, although the [ ] contrast also differs in manner. The dental/alveolar retroflex place d istinction, common to all of the chosen speech contrasts, allowed examination of factors (a) feature generalization and (b) multilingual benefit in three populations of subjects. Furthermore, selection of four different manners with same place contrast all ows for the examination of the learnability level of these speech contrasts. They can also provide new data testing the predictions of PAM in perceptual assimilation. Three subject groups were chosen for this study. Firstly, t he reason for choosing these p articular speech contrasts when looking at Bengali English (BE) language group, is primarily for the purpose of evaluating feature generalization. The speech sounds with retroflex feature [ ] are not found in Bengali However Bengali employ s the retroflex feature quite productively but only in the oral stop and flap manners of articulation. Therefore, if the retroflex feature get s generalized from existing phonetic categories to newly formed phonetic categories the evidence of any variabili ty in generalization of the feature in different manners will be reflected during the acquisition of these four new speech contrasts. Secondly, Spanish speakers altogether lack any experience with the dental/alveolar retroflex place distinction in phonemic or phonetic environments. However, as the participants of this study are bilingual (Spanish English), they are assumed to have phonetic experience similar to American English speakers, who have alveolar place distinction phonemically and sometimes have al lophonic experience with retroflex oral stops (See Polka 1991). No prior native language(s) experience with the retroflex feature makes the Spanish English bilingual group ideal to examine multilingual benefit factor This bilingual group, having signific ant

PAGE 34

34 improvement in identification of stimulus sets after training, if any, can be safely attributed to multilingual benefit The third population of subjects chose n for this study is a control group. Monolingual American English (AE) speakers were chosen i n order to control for the bilingualism factor. Also, American English speakers lack dental/alveolar retroflex place distinction at the phonemic level. However, retroflexion of alveolar stops in particular environmental contexts (e.g. [ ] 1991)) is sometimes found in the speech of American English speakers Organization T his chapter provided an introduction to the current study along with a review of previous literature regarding cross language speech perception as well as the posited fact ors A detailed discussion of the research questions evaluated by the study along with the relevance of the stimuli and sub ject populations we re presented. Chapter 2 gives a comprehensive description of the phonemic and phonetic inventory of relevant segme nts in the four languages utilized in the experiment. It also provides the experimental design of the perceptual training and testing paradigm used for the main experiment of the study. Chapter 3 presents experimental findings of the training experiment un dertaken in this study. Lastly, Chapter 4 summarizes the findings and discusses the implications and limitations of the current study.

PAGE 35

35 CHAPTER 2 METHODOLOGY Introduction The experimental design of this study is directed towards evaluating the hypotheses that (a) feature generalization and/or multilingual benefit may facilitate the learnability of non native speech contrasts by bilinguals and (b) assimilation patterns in the initial perception of non native speech contrasts can be used to predict subsequen t learnability in perception and accuracy in production. The study involve s the p erception training experiment which consists of learning of no vel non native speech contrasts. The next s ection describes the phonemic segments and their phonetic realizations in the four languages employed in the main experiment. The following s ection s describe the preparation and preliminary evaluation of the stimuli for the main experiment and discuss the recruitment of various language groups as subjects for the current stu dy. It also high light s the appropriateness of the choice of the three language groups in order to directly examine the research questions put forth in the study. Relevant literature review on the method of training and their underlying theoretical assumpti ons is also provided in subsequent section Towards the end of this chapter, t he adapted experimental design from the perception training p aradigm is described in detail. Description of languages The experiment outlined in this chapter entails the involvem ent of four languages. The perception training stimuli was obtained from Malayalam (target language). The subjects who underwent training were from three different language groups bilingual Bengali English and Spanish English speakers and monolingual Amer ica n English speakers With the description of each language, its distinctive segments are discussed with their phonetic allophones. For the proposed experiment, description of phonemic inventories alone would be insufficient to capture

PAGE 36

36 the initial percept ual exposure as well as the developmental acquisition of the novel non native speech contrasts. The comparison of all four languages in question regarding the retroflex feature is relevant to the assessment of factors such as feature generalization a nd/or multilingual benefit s ince exposure to phonemic or allophonic experience with the retroflex feature may be either beneficial or detrimental to the perception of a novel speech contrast ( Polka 1991 ; Strange 1995 ; Pruitt et al. 2006). The study looks at t he acquisition of Malayalam speech contrasts by Spanish English bilinguals, Bengali English bilinguals and English monolinguals. Malayalam A member of the Dravidian language family, Malayalam uses the retroflex feature in most of the manners it employs. Th is language was chosen to be the target language for the experiment since the retroflex place of articulation is highly productive in this language. Generally, retroflex sounds are taken as a specific place of articulation whereby a voiced retroflex stop i n one language would have similar articulatory characteristics in another language. Contrary to this opinion, Ladefoged and Bhaskararao (1983) revealed that retroflex speech sounds found in three languages, one Indo Aryan and two Dravidian, displayed a var ying degree of retroflexion. Hey claimed that the degree of retroflexion and which place of articulation the tongue blade touches could be compared with vowel heights in the way that it is possible to make a continuously changing set of sounds from an apic al dental to retroflex place of articulation. In their study, Tamil, a Dravidian language closely related to Malayalam showed a high degree of retroflexion whereas Hindi, an Indo Aryan language related to Bengali, showed least degree of retroflexion. Mala yalam, like Tamil, is assumed to have a high degree of retroflexion. For the purpose of the experiment, four speech contrasts were chosen from this language to serve as training stimuli. The common attribute in all of the four contrasts is the presence of the

PAGE 37

37 retroflex place feature in one of the segments in the contrast. The segments containing the retroflex feature are novel speech sounds for the listener groups. Table 2.1 shows the phonemic and phonetic segments of Malayalam (adapted from Asher, R.E. and Kumari T.C. (1997) ) Table 2 1. Phonemic c onsonant inventory of Malayalam (adapted from Asher, R.E. and Kumari T.C. (1997) Bilabial Labio dental Dental Alveolar Retroflex Palatal Velar Glottal Nasal m n n Approximant 1 j Trill/ Tap r Fricative s h Lateral approximant l Nasals Malayalam exhibits a rich inventory of nasal stops employing a six way place contrast starting with bilabial, dental, alveolar, retroflex, palatal up till the velar. Researchers agree on most of the phonemic distinctions among the different place contrast except for the dental alveolar place contrast. The dental alveolar place contrast has been discussed at length describing its phonological as well as phonetic distinction in the language (Kumari, 1972; Sreedhar, 1972; Kalackel, 1985; Raja, 1960; Mohanan and Mohanan, 1984; Asher and Kumari, 1997). languages, since dental and alveolar place contrasts have very subtle distinctions articulatorily and acoustically. It has been described in various ways on t he basis of its articulation a nd provided with phonemic pairs; t contact vs point contac t (Aiyar 1972), 1 There is a disagreement among researchers abo ut the phonetic properties of this sound. Asher and Kumari (1997) refers to it as a sublamino palatal approximant (discussed further in this chapter).

PAGE 38

38 apico laminal dental vs apico alveolar (Asher and Kumari 1997) or simply dental vs alveolar (Subramoniam 1973) have been used for this place contrast. to only one kind of distribution as geminates occurring medially. Due to this functionality, dental nasal and alveolar nasal are considered phonemic in Malayalam (Kumari, 1972; Kalackel, 1985; Asher and Kumari, 1997). However, Mohanan & Mohanan (1984) no ted that dental and alveolar nasals were perceived as separate in word initial as well as intervocalic positions in nonsense words by native Malayalam speakers, which perhaps implies that native speakers of Malayalam can distinctly perceive the dental and alveolar nasals even in distributions not found in Malayalam. On the other hand, apart from the geminate occurring medial position, the rest of the distributions in Malayalam for this contrast appear to be complementary to each other. The distribution for this contrast is described as [n ] occurring non finally and [n] occurring non initially (Kumari, 1972; Sreedhar, 1972). In addition, the orthographic script of the language does not distinguish between the dental and alveolar place contrast (Asher and Kum ari, 1997). For the purpose of this experiment, dental and alveolar nasals in Malayalam are assumed to maintain phonemic distinctions. Of particular interest for this experiment is the distribution of this place contrast in intervocalic position. Unlike t he dental nasal, the alveolar nasal occurs in intervocalic position (Raja, 1960; Kumari, 1972; Sreedhar, 1972; Kalackel, 1985; Asher and Kumari, 1997). The training stimuli employ VCV as the only syllabic position. Therefore, the speech contrast [n] [ ] results in being produced as alveolar nasal vs. retroflex nasal since intervocalically the nasal is realized as alveolar. This assumption does not conflict with Mohanan l and alveolar nasals distinctly in an intervocalic position. The perception of these segments in new

PAGE 39

39 phonetic contexts cannot be interpreted to mean that speakers can produce dental nasal intervocalically without any auditory prompt. Also, the reproductio n of segments correctly in term memory effect. Therefore, for the current experiment the contrast alveolar retroflex nasal was used as one of the training stimuli. Fricatives The voiceless frica tives in Malayalam do not have voiced counterparts. The fricatives in question, namely the lamino palato alveolar [ ] and retroflex fricative [ ] were introduced in Malayalam through Sanskrit loan words (Asher and Kumari, 1997). However, these fricatives i n to the language (Kalackel, 1985; Sreedhar, 1972 etc). The place contrast of palato alveolar retroflex fricative ( [ ] [ ]) was employed as one of the training stimuli from Malayalam. Laterals Laterals were chosen as one of the stimuli speech contrasts due to the fact that the lateral approximants in Malayalam exhibit a two way place contrast, alveolar and retroflex. This place contrast most appropriately measur es up to nasal or fricative place contrasts in terms of homogeneity of place of articulation as compared to the rhotic place manner contrast in the characteristic of Drav idian languages. This provides an opportunity to investigate the acquisition of this place co ntrast by Indo Aryan languages including Bengali which do not maintain alveolar retroflex lateral approxi mant place contrast

PAGE 40

40 Rhotics Most researchers talk of anot her lateral/ continuant sound which is considered unique to the D ravidian language family (orthographic symbol: ). However, there is disagreement to a large extent on its phonetic properties. It has been described in Malayalam as Voiced retroflexed palat al fri ci tivised lateral [ ] (Kumari 1972) Retroflex lateral fricative (no contact at the post alveolar region) [ ] (Sreedhar 1972) Voiced sublamino palatal approximant [ ] (Asher & Kumari 1997) Voiced apico prepalatal approximant [ ] (Kalackel 1985) Velar retroflex frictionless continuant [ ] (Raja 1960) Retroflex approximant [ ] or [ .] (Bright 1998 ; Krishnamurti 2003) In articulatory terms, Kalackel (1985) describes it as The apex of the tongue is grooved as if for pronouncing / /, but it is bent back so that the tip of the tongue is pointed at the root of the mouth where it is highest. The voiced sound which is thus produced is not actually a lateral (since the air escapes not over the sides of the tongue, but over the tip). Clearly, a detail ed phonetic description is required for this peculiar segment. For convenience, the term retroflex approximant along with the IPA symbol [ ] is used as reference to this particular speech sound in this study. This speech sound is contrasted with alveolar tap [ ] to create a paired contrast with rhotic properties. The case of alveolar tap retroflex approximant ([ ] [ ]) speech contrast is more complex than other alveolar/palato alveolar retroflex speech contrasts, in that the speech sounds contrast i n manner of articulation as well as place of articulation. This speech contrast is not homogenous with other speech contrasts chosen in terms of contrasting attributes between speech sounds. In all, four speech contrasts were chosen from Malayalam to exami ne the acquisition of novel non native speech contrasts by three listener groups. The novel speech sound in all of the contrasts is the one containing the retroflex place feature.

PAGE 41

41 The following subsections deal with the three listener languages, Bengali, S panish and American English. These languages will be discussed focusing on the presence or absence of retroflex features in their phonemic or phonetic segments. Bengali Bengali, an Indo Aryan language, displays productiveness of the retroflex feature in ma nners like plosives and flap. It was chosen as one of the listener languages. Unlike Malayalam, Bengali lacks the alveolar retroflex place contrast in manners like nasals, laterals and fricatives (Bhattacharya 1988 ; Ferguson and Choudhary 1960 ; Kostic and Das 1972 ; Tunga 1995 2 ; Ray et. al., 1966). These gaps in the consonant inventory make Bengali an appropriate choice of language to investigate the acquisition of alveolar retroflex place contrasts in these manners. In addition a comparing this langua ge group (Bengali English bilinguals) with the rest of the listener language groups allows us to investigate feature generalization of the retroflex feature. Bengali has a three way place contrast in nasal stops, namely bilabial, alveolar and velar. Unlike north western Indo Aryan languages (E.g Hindi), Bengali has lost the retroflex place feature in the nasal manner of articulation, although the Bengali script retains the orthographic symbol for the retroflex nasal. Therefore, one phoneme /n/ is used for t wo graphemes ( 2 Tunga (1995) noted that Midnapore dialect of Bengali (bordering the Oriya language area) uses the retro flex nasal [ ] more often than the dental nasal [n]. None of the subjects spoke this dialect.

PAGE 42

42 T he degree of retroflexion in Indo Aryan languages has been observed to be much less than that of Dravidian languages (Ladefoged and Bhaskararao 198 3). The retroflexion in Bengali is claimed to have a slighter degree of retroflexion than Malayalam (Chatterji 1921 ; Kostic and Das 1972 ; Haldar 1986 ; Tunga 1995 ). Table 2 2. Phonemic c onsonant inventory of Bengali Labial Post dental Alveolar Retrof lex Alveo palatal Velar Glottal Stops Fricatives s h Nasals m n ( ) Tap/flap Lateral approximant l Adapted from Bhattacharya, K. (1988) The fricatives in Bengali have similar story. Table 2 2 displays the phonemic inventory of Bengali. Phonemically, Bengali displays palat o alveolar [ ] and dental [s] fricative. The palatal fricative occurs in all syllabic positions as well as intervocalically as geminates. The dental fricative however, can only be realized in consonant clusters positioned word initially or occurs in loan words (Bhattacharya 1988 ; Fer guson and Choudhary 1960 ; Kostic & Das 1972 ; Ray et. al. 1966 ; Haldar 1986). Thus, the palatal p l ace contrast is the most productive in fricative manner of articulation. Bengali has lost the retroflex [ ] fricative originally derived from Sanskrit. How ever, as with the nasals, the Bengali script retains the orthographic symbol for the retroflex fricative. There are three graphemes (

PAGE 43

43 Bengali maintains a single place in the case of lateral approximants, alveolar /l/ (orthogr aphic symbol: It lacks the retroflex place of contrast in laterals as a distinctive segment. However, Bengali speakers do experience retroflex lateral [ ] as an allophone of /l/ in certain environments. Alveolar laterals preceding retroflex consonants undergo assimilation and are realized as retroflex lateral approximants, especially in consonant sequences like / / or / / (Bhattacharya 1988 ; Ferguson and Choudhary 1960 3 ; Tunga 1995). Rhotics in Bengali have a two way place contrast, namely alveolar / / and retroflex / / flap. The alveolar flap occurs in medial and word final positions. Alveolar trill [ r ] is realized as an allophone of this phoneme when it occurs word initially. Ferguson & Choudhary (1960) observed that sometimes an alveolar flap occ urring after a labial or retroflex stop gets realized as an approximant (like English [ ]). Similarly, the retroflex flap when occurring in final position or followed by a consonant sometimes is realized as a prolonged retroflex continuant. Therefore, Beng ali speakers may have allophonic experience of retroflex approximant as well. Nevertheless, none of the Malayalam retroflex speech sounds being tested for acquisition occur phonemically in Bengali. Spanish Spanish, one of the Romance languages, uses the pa latal place of articulation extensively. Most Spanish dialects allow manners of articulation like nasal, fricatives, approximants and laterals at the palatal place of articulation (Hammond 2001). This language was chosen as another listener language invol ving the Spanish English bilingual group, based on the 3 Ferguson & Choudhary (1960) also claims that a sequence like/ / (where alveolar lateral follows a retroflex consonant, both consonants result in a long retrofl ex lateral [ ].

PAGE 44

44 nonexistence of the retroflex feature in its phonemic as well as phonetic segments. The lack of experience with retroflexes makes the Spanish speaking subjects ideal to test the effects of meta linguis tic awareness since any kind of linguistic experience with retroflex is controlled in this language group. Table 2.3 shows phonemic consonant inventory of Spanish for the relevant manners of articulation, namely rhotics, nasals, approximants and lateral ap proximants. Table 2 3. Phonemic consonant inventory of Spanish (Only manners that are relevant for the study are displayed here.) Bilabial Labio dental Dental Alveolar Post alveolar Palatal Velar Trill Tap Nasal Fricative [ ] [ ] [ ] Approximant (labio) Lateral approximant *Adopted from: Hammond (2001) Rhotics in Spanish are considered another unique attribute of Romance languages. The alveolar trill is phonemically contrasted with the alveo lar tap (Lindau 1985). The alveolar trill is not found as a phonemic sound in either English or Bengali. In the case of the alveolar tap, while being a phonemic consonant in Bengali, it occurs only in phonetic relationship with alveolar /t/ and /d/ stops in American English (Roach 1983 ; Olive et. al., 1993; Ladefoged and Maddieson 1996). Nasals as well as lateral approximants in Spanish have a two way place contrast described as alveolar and palatal. This contrast is prevalent in most of the South America n dialects of Spanish In the case of fricatives, most of the Spanish dialects maintain the alveolar place contrast in sibilants and do not have an equivalent of palatal or palato alveolar fricative ; the exception being Argentinean Spanish where in the fri cative manner a palatal allophone also occurs along with the alveolar one.

PAGE 45

45 There appears to be complete lack of retroflex features in the sound inventory of Spanish whether phonemic or phonetic. Therefore, the acquisition of retroflex sounds or specificall y the alveolar retroflex place contrasts from Malayalam may pose a degree of difficulty, attributed to acquisition of new non native speech sounds, for the speakers of Spanish. American English American English was included as the third listener language f or the experiment. This listener group (monolingual American English group) acted as the control group for the experiment. English employs the alveolar place of articulation extensively to create phonemes in different manners of articulation. Table 2.4 sho ws the phonemic consonant inventory of American English of the relevant manners of articulation. Table 2 4. Phonemic consonant inventory of American English (Only manners that are relevant for the study are displayed here.) Bilabial Labio dental Dental A lveolar Post alveolar Palatal Velar glottal Nasal Fricative Approximant (labio) Lateral approximant *Adapted from: Ladefoged (1999) The lateral approximant in American English employs only the a lveolar place. The phoneme /l/ has a velarized lateral approximant allophone, syllable finally. Nasals in American English have a three way place contrast namely, bilabial, alveolar and velar. The post alveolar fricative, similar to palato alveolar fricati ve in Malayalam, does exist in American English. However, the langauge lacks any existence of retroflex fricative. The alveolar tap, while being a phonemic consonant in Malayalam, occurs only in a phonetic relationship with alveolar /t/ and /d/ stops in Am erican English (Ladefoged and Maddieson 1996). In the case of retroflex feature exposure, unlike Spanish, English speakers

PAGE 46

46 may have some phonetic experience with the retroflex feature in incidental contexts. Polka (1991) noted that English speakers may ha ve some experience of retroflex stops at the phonetic level. The alveolar stops /t/ and /d/ may be produced as retroflex stops / / and / /when the alveolar stops precede / However, as noted by Polka (1991) this is not a predictable phonological process and may vary on a dialectal or individual level. Also, it should be noted that the consonants in question (/l, n, r, sh/) do not precede the rhotic approximant in syllable initial position in American English. Therefore, AE speakers do not have the experience of their becoming retroflexed as is the case with /t, d/. Stimulus Materials The novel non native speech contrasts for the study were taken from Malayalam. It was chosen as the target language as it has a rich inventory of consonants employing the retroflex feature extensively in various manners. This allows a flexible choice of speech contrasts as per the requirements of the study (as discussed below) The speech contrasts chosen for the study are: (1) [ ] ( alveo lar versus retroflex lateral approximant) (2) [ ] (dental /alveolar versus retroflex nasal) (3) [ ] (palato alveolar versus retroflex fricative) (4) [ ] (alveolar tap versus retroflex approximant) These speech contrasts differ primarily in place of articulation, although the [ ] contrast also differs in manner. The reason for choosing these particular speech contrasts when looking at the Bengali English (BE) language group, is primarily for the purpose of evaluating feature generaliza tion. One of the speech sounds in each speech contrast from Malayalam is marked with the retroflex feature [ ] These sounds are not found in Bengali which employs the retroflex feature but only in the oral stop and flap manners of articulati on.

PAGE 47

47 Therefore, if the retroflex feature get s generalized from existing phonetic categories to newly formed phonetic categories the evidence of any variability in generalization of the feature in different manners will be reflected during the acquisition o f these four new speech contrasts. In case of the monolingual language group, American English (AE) speakers do not have experience with retroflex sounds phonemically. However, retroflexion of alveolar stops in particular environmental contexts (e.g. [ is found in the speech of American English speakers Spanish English (SE) bilinguals have no experience with the retroflex feature in phonemic or phonetic environments when using Spanish. However, being bilingual, they are assumed t o have phonetic experience similar to American English speakers. Having n o phonemic exposure to retroflexion makes the Spanish English bilingual group ideal to examine the multilingual benefit The materials were recorded as natural stimuli by Malayalam ta lkers producing each of the eight speech sounds from the four speech contrasts discussed above. Each speech sound was embedded between two identical vowels (vowel contexts: [a, i, u]) resulting in a set of non words. Six repetitions of each consonant with each vowel context were recorded in order to use the best possible tokens for the stimuli Recording for the stimuli was done in India (Chennai, Chandigarh, Bangalore, Cochin, and Ahmadabad) Currently, all speakers that were recorded have been residing ou t of the state of Kerala. Ten native speakers of Malayalam were recorded (See figure 2 1 for their region distribution) However, due to errors in elicitation, two speakers ( M03 and M05 ) were not considered for this study. Detailed information of the Malay alam speakers is provided in table 2 5

PAGE 48

48 Table 2 5. Talker information; all talkers are native speakers of Malayalam and born in the state The data were recorded and digitized at 44.1 KHz on a Marantz Digital recorder and transferred directly onto a personal computer. The recordings took place in a quiet room with no significant reverberation or background noise. The subjects were provided with the non words (in Malayalam script ) on the computer with an inter trial interval (ITI) of 2.5 second s. The talkers were asked to read through the list before the recording in order to be familiar with the non word materials The data w ere segmented using PRAAT speech analysis software. Talker Code Gender Age Years Other Languages known M01 F 28 4 Kannada English M02 F 37 1 Hindi English M04 M 45 15 Hindi Punjabi English M06 F 36 9 Hindi M07 M 28 1 English Tamil M08 M 29 5 Hindi English M09 M 32 8 Hindi English M10 M 29 5 Hindi English

PAGE 49

49 Figure 2 1. A map of the state of Kerala with cities of origin for the Malayalam speakers. A smaller map of India on the left hand bottom corner shows the location of the state of Kerala. Stimulus Material Evaluation Before developing the stimulus m aterial for the main study, the recorded data from the Malayalam talkers was evaluated. Native Malayalam listener ratings were elicited for tokens from eight talkers. Four native speakers of Malayalam evaluated the intelligibility of the recorded stimuli t hrough a consonant identification task. Tokens that were identified consistently as the target sound 83% 100% were selected for the stimuli. T hree out of the selected tokens closest in duration and mean fundamental frequency, were selected for each cons onant with all three vowel contexts, making a total of 72 stimuli (8 consonants x 3 vowel contexts x 3 tokens) per talker.

PAGE 50

50 Stimuli Evaluation: Preliminary Experiment C onducting the main experiment would have been unproductive for the purpose of evaluating the role of assimilation if ceiling or floor effects were observed for any of the speech contrasts in the monolingual language group (AE) Therefore, to avoid this possibility, a preliminary experiment was conducted to examine the discrimination level of t hese contrasts by monolingual American English speakers since the AE monolinguals acted as the control group in the main experiment, against which the results of the bilingual groups are to be compared. Thus, it is considered reasonable to test the difficu lty level of the stimuli with the control group. The goal of the preliminary experiment was t o examine the level of discriminability of the novel non native speech contrasts by American English speakers. Eleven monolingual American English subjects were re cruited for this experiment. All were undergraduates at the University of Florida. They were compensated with extra credit or paid cash ($5/40 minutes). None of them reported average proficiency in any language other than American English. For stimuli, t wo tokens of each non word (8 consonants x 3 vowel contexts) per talker were used Data from three talkers was used. In all, there were 1152 stimuli pairing up to make 576 trials (16 trials per speech contrast different trials, identical trials and physica lly different trials). Subjects were given a categorical AX discrimination task (Please refer to the section later in this chapter for a description of an AX discrimination task.) Because physically different tokens of the stimuli are used, the discrimina tion task has to be administered as a categorical task. Naturally produced stimuli may have subtle irrelevant acoustic differences which may level processing. Therefore, the subjects are asked to decide wheth er both stimuli sound like they are from same category or different

PAGE 51

51 category. For this task, t he inter stimulus interval (ISI) was provided as 1 sec and inter trial interval (ITI) was 4 sec. The results of the preliminary experiment were as follows. Each s the AX test was converted to a d score (scale of 0 to 4) Table 2.6 shows the results of the discrimination task for the American English speakers. Table 2 6. Result of the preliminary AX discrimination task across three Malayalam speakers. Mean D prime values averaged across contrasts as well as across vowels. [ ] [ ] [ ] [ ] Average across contrasts [a] 1.09 2.55 2.99 2.70 2.33 [i] 0.94 1.11 1.69 2.28 1.51 [u] 1.37 1.25 2.89 1.09 1.65 Average across vowels 1.13 1. 64 2.53 2.02 The [ ] contrast averaging over all three vowel context was the most discriminable of the four speech contrasts (d value 2. 5 ). Overall, the [ l ] speech contrast was least discriminable (1 1 ) followed by the nasal contrast [ n ] (1 6 ) and fricative contrast [ ] with d value at ( 2.02 ). A mong the vowel contexts, the [i] context was least discriminable (1.5) closely followed by [u] (1. 6) whereas [a] was the most discriminable (2. 3 ). Across the four speech contrasts, the [i] vowel context was the least discriminable (with the exception of fricative [ ] contrast where [u] was the least disciminable) and [a] was the most discriminable. No ceiling or floor effects were found for any of the non native speech contrasts. Stimulus Prep aration Based on the results of the preliminary experiment as well as native listener ratings, the stimulus materials for the perception training experiment were selected. For each talker, three physically different tokens per consonant and vowel context ( 8 consonants x 3 vowel contexts x

PAGE 52

52 3 tokens ) were chosen. The tokens of one talker (M08) were used solely for the familiarization task T alker M04 was used for pretest and posttest as well as discrimination test and perceptual assimilation test since all o f these tests have to be identical in stimuli and procedure in order to be directly compared. M04 was also included during training since the testing (pretest and posttest) is on a trained talker. M07 was used for the generalization test only. This would b e the novel talker, whose voice was not exposed to the subjects anytime during the pretest or training phase. D ata of six talkers (M01, M02, M04, M06, M09, M10) was used for training sessions. Familiarization phase: one talker (8 consonants x 3 vowels x 2 tokens x 2 repetitions) 96 trials in all. Pretest/posttest phase o Pre test/post test: one talker (8 consonants x 3 vowels x 2 tokens x 6 repetitions) 288 trials in all. o Discrimination test: one talker (8 consonants x 3 vowels x 2 tokens x 4 orders) 192 tri als in all. o Perceptual assimilation test: one talker (8 consonants x 1 vowel x 2 tokens x 5 repetitions) 80 trials in all. o Production pretest/posttest: one talker (8 consonants x 3 vowels x 1 token x 6 repetitions) 144 trials in all. o Generalization test: o ne talker (8 consonants x 3 vowels x 2 tokens x 6 repetitions) 288 trials in all. Training phase: two talkers per session (8 consonants x 3 vowels x 3 tokens x 2 repetitions) 288 trials per session. Over a total of six training sessions, data from six talk ers is used. Sound symbol a ssociation i ssue : The issue of sound symbol association was foreseen in the design of the training experiment. Since the subjects (all three language groups) are not familiar with Malayalam orthography or vocabulary, it w ould be difficult to train them without sound symbol association. This kind of issue has not arisen in the previous studies (Logan et al., 1991 learning of /r/ /l/ distinction by Japanese learners of English) as the subjects were familiar

PAGE 53

53 with the vocabulary and alphabets of the target language. In the current study, t he question of orthography associated with the 8 consonants was resolved by using arbitrary symbols for each consonant. The arbitrariness of the symbols being associated with the sounds is inevitabl e since these are novel non native sounds for the listeners. A number of alternate methods of representing the sounds were considered before appropriate symbols were finalized. Firstly, using the original Malayalam script or IPA symbols with the sounds w ou ld have take n the subjects much more time and effort for the sound and symbol association resulting in decreased focus o n the perception of sounds Secondly, purely arbitrary geometric shapes as symbols were d ismissed as potential candidates for the same r eason as above. In addition to that, the confounding possibility of learners perceiving the sounds correctly but clicking on the wrong geometric shape could be detrimental to the validity of the results. The third alternative was using the English alphabet s as representative symbols. The option was considered as the subjects in all the language groups were familiar with English alphabets and their corresponding English consonants. The small and upper case of the alphabet ] [ ] sound distinction so that whenever subjects hear a nasal, there would n o t be the basic confusion of finding the correct symbol out of the 8 symbols, as would be the case with purely arbitrary representations. One caveat with this type of representation co uld be that the subjects may not easily learn the distinction between the small and upper case letters representing the alveolar retroflex distinction. Therefore, to avoid this possibility, all dental/alveolar sounds in the four sets of contrasts that were presumed to be assimilated to their L1/L2 category were assigned the small case letters. The retroflex sounds in all the four sets of contrasts were assigned the corresponding upper case letters. In addition to this, during familiarization process, the so unds in each set were auditorily prompted (along with their

PAGE 54

54 corresponding symbol) next to each other. The subjects were told that the sounds they will hear are not English sounds but from a language not known to them. After careful consideration, the ortho graphic representation of the sounds was finalized as small and upper case English alphabets. Hence, the study uses the following symbols in training: l= dental lateral approximant; L=retroflex lateral approximant; n=dental nasal; N=retroflex nasal; sh=pal ato alveolar fricative; S=retroflex fricative; r=alveolar tap and R=retroflex approximant. These arbitrary symbols were familiar to all three language groups and their efficacy was verified in pilot testing. Subjects The study employed a diverse range of c hallenging non native consonant contrasts from Malayalam with three populations of listeners, Bengali English (BE) bilingual group Spanish English (SE) bilingual group American English (AE) monolingual group Twenty (20) participants were recruited for eac h group from the University of Florida campus. At the time of the experiment, the participants, ages ranging from 18 to 35, were enrolled in undergraduate or graduate courses at the university. For all subjects the level of proficiency was self reported t hrough a language background questionnaire. Self reported language proficiency may not be an accurate measure for screening of potential subjects for the study. Therefore, certain precautionary measures were taken while recruiting the subjects for the thre e language groups. In the case of American English monolinguals, the experimenter asked for specific information regarding the knowledge of other languages or extent of exposure to other languages apart from English ( L1 ) The preliminary (introductory) ema il sent out to the interested individuals has been provided in Appendix A If

PAGE 55

55 the candidate had taken Spanish language classes in middle/ high school, which was mostly the case, a brief test of spoken Spanish was conducted to assess the level of L1 transfe r from English The procedure and content of the spoken Spanish test has been explained in Appendix A In the case of Bilingual Spanish English subject recruitment, in addition to conducting the language background questionnaire, the experimenter assessed the level of proficiency in spoken American English. These individuals were also asked to take the spoken Spanish test. The screening for bilingual Bengali English subjects involved assessing their level of spoken proficiency in English by observing any no ticeable signs of L1 transfer in consonantal segments during conversation. More specifically, the Bengali English participants were screened for any exposure to Dravidian languages and if present then to what extent Candidates with eve n minimal awareness of retroflex sounds were not recruited for this study. During the screening process, bilinguals, preferably early bilinguals with an age of acquisition ( AOA ) for L2 of fewer than 1 2 yrs were considered as potential subjects. If age of acquisition (AOA) fo r L2 was more than 12 yrs, then candidates with self reported proficiency level of L2 as HIGH were considered as potential subjects. Specifically for Bengali English subjects, the English refers to Indian English since most of the student population is gra duate reported and also assessed by the experimenter. For the monolingual subjects, self reporting of any language exposure was required including conversational abili ty in another language or passive exposure to another language. Subjects with very minimal knowledge of another language (mostly Spanish from High School language classes< 2 years) were considered in the monolingual American English language group. It was almost impossible to find pure monolingual subjects with no exposure to another

PAGE 56

56 language especially in the age group of 18 35 years due to the educational curriculum in North American Schools. None of the participants had any prior exposure to Malayalam. I n addition to this, subjects were asked in the screening questionnaire (Appendix A) of any hearing or speech problem that they may have had in the past. All subjects reported normal hearing with no hearing or speech impairment history. No preliminary clini cal screening for normal hearing was conducted. Therefore, the information collected on normal hearing was self reported. Bengali English Bilinguals The Bengali English language group was chosen for their phonemic experience with the relevant place feature (retroflex in oral stop and flap manners of articulation), though Bengali and English lack any direct correspondents to the Malayalam contrasts under study. The results from the BE language group when compared with Spanish English bilinguals were expected to provide evidence for the feature generalization factor in th e learning of a third language. Twenty (20) speakers of Bengali with English as L2 participated in the study. The mean age of participants in this group was 28 years All subjects were from Ko lkata, the capital of West Bengal state with the exception of two subjects. BE16 and BE17 are from Barddhaman, a southern district in West Bengal, close to Kolkata. They speak the standard colloquial Bengali, similar to what is spoken in Kolkata. Most of t he subjects have acquired a working knowledge of Hindi through entertainment media, but only four subjects, BE03, BE05, BE11and BE12, have had formal education in Hindi. Some subjects have been exposed to languages like Assamese, Oriya, Urdu, Nepali and Te lugu through the surrounding environmental context from a few months up to two years. However, even low proficiency in comprehension or speaking in these languages was not reported. Therefore, any influence from the above mentioned languages can be ruled o ut. The details of all the subjects have been provided in the appendix A

PAGE 57

57 Spanish English Bilinguals The Spanish English (SE) bilingual group was chosen for their lack of linguistic experience with retroflex place feature. The results from SE language grou p when compared with other groups were predicted to provide evidence for effects of multilingual benefit in bilinguals. Twenty (20) speakers of Spanish with English as their L2 or another L1, participated in this study. The mean age of subjects within this language group was 20 years All subjects have lived in North America all or most of their lives. In cases where subjects were born outside of North America, they had shifted to North America within the first 6 years of their lives. Therefore, these bilin guals acquired Spanish at home as their first language and picked up American English either simultaneously with Spanish or during preschool years (age 3 to 6). The dialects of Spanish spoken by these subjects are of Central American o rigin with the except ion of one subject who spoke Argentinean Spanish Figure 2 2 shows the origins of the Spanish dialects of the SE bilingual subjects. The table providing detailed information on individual subjects is given in appendix A.

PAGE 58

58 Figure 2 2. An outline map of Me xico, Central America and South America showing the regional distribution of the subjects. American English Monolinguals American English speakers were recruited for the monolingual language group acting as a control group for the experiment. The subjects (20.2 years as the average age within this language group) acquired American English as their first language. None of the subjects reported more than minimal proficiency in any language other than American English. Since most of the

PAGE 59

59 students had lived mos t of their lives in Florida, many of them had to take Spanish as a required language class. This factor could not be controlled for therefore, subjects were asked to read out two brief tongue twister phrases in Spanish. This was done to determine if the su bject had actually acquired any perceptual categories corresponding to Spanish language. Only those participants were recruited as subjects whose speech did not provide any evidence of Spanish categories but in fact displayed transfer from American English The Spanish phrases as well as detailed information on individual subjects are provided in Appendix A Experimental Design Perception Training Method The present experiment used a perception training method with identification paradigm containing high va riability stimulus (High variability perception training). This training emerged as an effective training method in learning non native speech contrasts through the academic disagreement between two schools of thought regarding the mental representation of phonetic sounds in the minds of L2 learners: the prototype model (Strange and Dittmann 1984 ; Jamieson and Morosan 1986 1989 ; Kuhl 1991, 1992) and the exemplar based model (Jamieson and Morosan 1986 ; Logan et al. 1991 ; Lively et al. 1993 1994 ; Brad low et al. 1997). In order to comprehend the reason for adopting High variability perception training method for this study, it is imperative to discuss the theoretical assumptions underlying these two models of long term memory representation. The earlie r school of thought, the prototype model, claimed that the phoneti c sounds acquired by a listener were represented as context invariant units of sounds, the content of which was the ideal example of that phonetic sound, otherwise named as the prototype, or in a phonologically driven view, phoneme. Every representation of a phonetic category in the mind of an L2 learner was assumed to have foci the most representative member of the category

PAGE 60

60 (Kuhl 1991, 1992). The foci was expected to have more stable repre sentations in long term memory, since only linguistically relevant information was stored in that prototype and could overlap concrete natural members of that category when perceived. However, empirical testing of this training type failed to prove effecti ve in developing long term memory representation since the subjects in these studies failed to generalize to new phonetic environments. A limitation in the notion of prototypical representation was that it was assumed that only linguistically relevant info rmation will be extracted to form the representation of the phonetic category. However, by postulating this, the model assumed that other detailed information in those stimuli out. This detailed information was established to be of relevance in the representation of phonetic categories by subsequent empirical results (Logan et a l. 1991 ; Lively et al. 1993). The results from Logan et al. (1991, 1993), Lively et al. (1994) and B radlow et al. (1997) clearly indicated that for generalization to occur, detailed information like talker variability and phonetic environments are as relevant as other linguistically significant information. The perceptual learning and development of stab le representations of phonetic categories result from extensive stimulus variability in natural stimuli. This led to the formulation of the exemplar model of phonetic representation, extending the exemplar based model of categorization (Nosofsky 1986, 198 7 ; Kruschke 1992) based on the selective attention weighing mechanism. Multidimentional representations are stored in the memory during training while selective attention weighs the importance of various stimulus dimensions. Changes in selective attention structure resulting in phonetic categories being less similar to other members as dimensions are

PAGE 61

61 stretched and more similar to other members when dimensions ar e shrunk (as cited in Lively et al. 1993) The choice of training method for the current research is clear from the above discussion. The present experiment used a perception training method with identification paradigm containing high variability stimulu s (High variability phonetic training). The choice of perception over production or pronunciation training is also based on the debate of mental representation of phonetic categories in long term memory. Since production/pronunciation training is based on providing the learners with the ideal or best exemplar of the phonetic category, the input lacks any variability at all (Dalby and Kewley Port 1999 ; Akahane Yamada et al. 1996; Neri et al. 2002 etc). Procedure A pretest training posttest design was empl oyed adapting the procedure used by Lively et al. (1993, 1994). The effect of training was measured by comparing the performance in pretest and posttest tasks A consonant identification procedure was used throughout the training phase of the experiment. Apart from the consonant identification task, an AX discrimination task and perceptual assimilation task were also employed during the pretest and posttest phase of the experiment. A minimum of seven hour gap was used between sessions to avoid any confound ing effects of working memory influencing the perceptual performance before and after training and also to prevent any adverse effects of fatigue during consecutive training sessions. The term working memory, in this case, refers to immediate retrieval of information stored in the short term memory as a result of preceding training. On the contrary the results would be more valid This arbitrary period of seven hours as t he minimum gap between two consecutive train sessions was considered assuming that the routine task of the individuals would negate any working memory

PAGE 62

62 effects. All training and testing took place in a sound attenuated room equipped with individual computer stations (contains a keyboard, a CRT monitor and headphones) for subjects. Stimulus materials were presented to subjects over a set of calibrated head phones. The software collected individual responses during all phases of the experiment. All subjects we re tested and trained individually. Familiarization Subjects were familiarized with the symbols associated with each consonant. The talker used for this task hyper articulated the sounds during recording and, thus, was considered ideal for the familiarizat ion task. The stimuli consist ed of one talker (8 c onsonants x 3 vowel contexts x 2 tokens x 2 repetitions) In all, there were 96 trials and the ITI was 2 sec onds The sound file was played while the symbol was displayed on the computer screen. The subject s were asked to pay attention to both the speech stimulus and its corresponding symbol. They were told that the words are from a language not known to them and the arbitrary symbols represent different sounds. The non words contrasting in place (eg. [ili] vs [iLi] ) were presented within same vowel context alternating between the two words for maximum impact of familiarization task. Pretest During the pretest phase, the perceptual performance of the non native stimuli was determined at the nave listener lev el through various tasks Pre testing of perception (consonant identification task) was incorporated to test the level of accuracy in identification of the target sounds. The stimuli for the pretest (identical task also given as posttest ) consist ed of one talker ( 8 consonants x 3 vowel contexts x 2 tokens x 6 repetitions) whose data is a lso used for the trai ning (Trained on talker). In all there were 288 trials. The pretest consist ed of a consonant identification task presenting the randomized stimuli over the h eadphones. The sound file was played an d the response options beca me available for the subject to click on. The response

PAGE 63

63 options provided were all of the eight symbols corresponding to the eight non native sounds in question, thus technically making 12.5% as chance level though no confusion between the manners is expected since all L1s have dental/alveolar nasals, laterals and rhotics The subject ha d the option to change her /his response before going on to the next trial. Once the next trial button w as pressed, the answer was recorded by the software and no changes could be made (same for all tasks). The task was self paced and therefore no ITI was assigned. No correct answer feedback was provided during pretest. A categorical AX discrimination task w as used to measure listener sensitivity to non native speech contrasts at the nave exposure level (pretest). The same task was administered during the posttest phase so as to measure the perceptual performance on the same stimuli after training. An AX dis stimulus set. Each trial consists of a set of two stimuli (either category different or category same). Subjects are required to decide if the stimulus X (the second stimulus t hey hear) is the same o r different than stimulus A (the first of the stimuli set) The decision of similarity or difference is required to be made on the basis of perceptual categories (phonemic sound, word etc). Two types of trials can be used in such dis crimination task: physical identity trials and categorical trials. The physical identity trials only pair up the same token for same trials so that the trial represents the same exemplar of the speech sound. This provides insight into the perception of pu rely physical properties of the sounds. On the other hand, categorical trials, along with using physical identity trials, also employ the physically different tokens of the same speech sounds as pairs for the same trials. This provides insight into the per ceptual processing at the phonemic level. For the discrimination task in this study, categorical trials were employed.

PAGE 64

64 Twelve possible trials per speech contrast are usually employed. An example of these possible trials is given in figure 2 3. In this test t he two same trials are repeated in order to maintain equal number of same and different trials and therefore the number of same trials become four making a whole set of 16 trials per speech contrast Inter stimulus interval (ISI), the time gap between the presentations of the stimuli in a single trial, is also important in defining the difficulty level of the task. The longer the ISI, the more difficult it is for the subject to rely the response only on working memory. Thus, phonemic level processing co mes into play and the longer ISI tests the perception at a categorical level. Same trials A1 A2 X1 X2 A2 A1 X2 X1 Different trials A1 X1 X1 A1 A1 X2 X2 A1 A2 X1 X1 A2 A2 X2 X2 A2 Fig ure 2 3. Different trials possible in an AX discr imination task using two tokens of each speech sounds. Another task known as the p erceptual assimilation task was administered during the pretest phase to examine the similarity of the speech sounds presented as stimuli when compared with tive speech sounds. This test provides an insight into listener intuitions of classifying the non native speech sounds according to the already established phonetic categories of the native language(s). In this study, the listeners we re asked to listen to each stimulus sound (provided with only one vowel context) and write down the perceptually closest speech sound (in the native language orthography) in her native language(s). In case of bilingual groups, the subjects were allowed to ch oose the response (t he closest sound) from both L1 and L2. They were asked to write down the alphabet(s) from their native language corresponding to th e sound heard and specify the language (especially in the case of Spanish English bilinguals since the

PAGE 65

65 writing system of both English and Spanish overlaps substantially). This assimilation task was an open set test, where the listener was not bound by a few response options to choose from but was allowed to choose from her past native linguistic experience. Along with that, for each trial, listeners we re asked to provide a category goodness rating of how similar the native speech sound is to the newly heard stimulus sound. The goodness rating scale consisted of 1 to 7 levels ry (modal response for the token, according to the listener) Listeners were able to specify the degree of difference between the sounds by circling a number between 1 and 7. The same task was administered at the both the pretest as well as the posttest phase of the experiment in order to evaluate any changes in as similation patterns as an indication of developing new phonetic categories. In all, the pretest phase consisted of a consonant identification test, an AX discrimination test and a perceptual assimilation task. Training Subjects began the 6 session training phase after the pretest phase. The training procedure used the same consonant identification task. The difference between the pretest task and the training task is that feedback is given on every trial during the training sessions. If the response was cor was replayed along with the correct symbol for reinfor cement of the association. If the response was Incorrect!!! Please listen to it again w as displayed A fter 500 ms interval the sound file was replayed and the correct symbol was displayed for reinforcement. The subjects were not required to respond during repetition. The task was self paced with no ITI between trials. Data gathered from six talkers was used as stimuli for the training portion of the

PAGE 66

66 experiment. Subjects hear d only two talkers per training session. The sequence of presenting talkers remained the same throughout training. Therefore, subjects hear d each talker two times during t he 6 day training phase. The stimuli contain ed (8 consonants x 3 vowel contexts x 3 tokens x 2 repetitions) per talker. In all, there were 288 trials (2 talkers per session). The task was paused for 60 seconds after one block of trials (stimuli of one talk er). The s ubjects were allowed to take break s for longer time if required Posttest A final posttest was given at the end of the training phase. The posttest consist ed of the consonant identification task discrimination task and perceptual assimilation ta sk ( all three identical to pretest ). Another similar identification task was administered but with untrained stimuli and one talker (novel talker) to measure the generalization ( generalization test ) of the training to novel stimuli with new talker voice in formation. The design of the generalization test was identical to the consonant identificati on test except that the stimuli w ere taken from a novel untrained talker. Task Order Presentation The order of presentation for the various tasks during the pretest phase was kept same for all the subjects across language groups. Once the familiarization task was over, the pretest phase began with the consonant identification pre training test as the initial task. It was followed by the AX discrimination test. After a brief rest for five minutes, the subjects were given the perceptual assimilation task. The presentation of the stimuli in each of the tests was identical for all the subjects. The consonant identification test, AX discrimination test and the perceptual a ssimilation task that were conducted at the pretest level were repeated at the posttest level of the experiment as well. In addition to that, a generalization test was also conducted during the posttest phase

PAGE 67

67 The presentation order of these tasks was as f ollows: consonant identification post training test followed by the AX discrimination test and the perceptual assimilation task. Towards the end, the generalization test, where perceptual performance on stimuli from novel talker was assessed, was conducted Data Analysis The results of the pretest and posttest phase were analyzed across contrast and vocalic context as well as across language groups. The mean percentage scores for the consonant identification tasks and d prime scores (discussed further in de tail) for the discrimination task s were obtained. Results from the generalization test were analyzed across groups and individually when compared with posttest results. ANOVAs were conducted separately for each contrast to test the hypotheses under conside ration. The calculation of d prime scores and perceptual assimilation types are discussed here with were then converted to a binary code, termed for this study as signal present. Hit rate, which means that the subject correctly identifies the distinction within a contrast, was calculated by averaging the signal pre subject has marked all responses the same regardless of the trial type (either same or different), brings out such discrepancies in the result and helps in accurately obtaining the d prime score. Based on these, the d prime score was then calculated by subtracting the NORMINV value of false alarm rate from that of the hit rate. NORMINV for mula returns the inverse of the normal cumulative distribution for the specified mean and standard deviation. Once the d prime values were obtained for the pretest discrimination test as well as the posttest discrimination test, a

PAGE 68

68 difference score (posttes t value pretest value) was obtained and later averaged vowel context per contrast for each subject. These difference scores were submitted to the statistical tests for the discrimination results. The calculation of perceptual assimilation types from the da ta collected during the PA task was as follows. The responses written by the subjects were recorded in excel files. To arrive at the modal response for each sound, a matrix was developed which calculated the percentage of a particular response for that spe ech sound. Once the modal response for each of the sounds was reached at, the general assimilation types were obtained. The resultant general assimilation types for each contrast (pre as well posttest values) were either within category (WC) or between cat egories (BC). A contrast was BC if the modal responses for the sounds in that contrast were different. On the other hand, a contrast was WC if the modal responses were same and the overlap between them was less than or equal to 0.1 (m axi mum). For the betwe en category assimilation types, if the overlap was more than or equal to 0.9, it was termed as two category assimilation (TC) since the overlap of modal responses was minimum. If the percentage count for modal responses for both speech sounds within a cont rast was less than 0.9, then it was termed as both uncategorizable assimilation types (UU) otherwise it was termed as uncategorizable categorizable assimilation type (UC). In the case of within category assimilation types, the ratings for the speech sounds in that contrast were submitted to TTest (statistical measure) in order to examine any significance difference between the goodness ratings given for the modal responses corresponding to each speech sound. If significance was found between the ratings gi ven, the contrast was termed as category goodness assimilation type (CG), otherwise single category (SC). (Please refer back to Chapter 1 for detailed description of the various assimilation types under Perceptual Assimilation Model.)

PAGE 69

69 Summary In this chapt er, the methodology of the main experiment was provided. Starting with the description of the target language (Malayalam) and three listener languages (Bengali, Spanish and American English), the chapter provides details of stimuli preparation and evaluati on as well as issues related to stimuli presentation. Information on subject groups was given including the explanation to the appropriateness of the choice of the subject groups. A review of the perception training paradigm involving discussion of theore tical issues in mental representation of perceptual categories is considered essential for the understanding of the perception training method chosen. Finally, a detailed procedural description of the exp eriment was provided

PAGE 70

70 CHAPTER 3 RESULTS Introducti on For this study, three goals were put forth in chapter one: to investigate the factor of multilingual benefit in cross language speech perception; to examine the role of feature generalization in facilitating the acquisition of non native speech contrast s by trained / experienced listeners; and to investigate changes in assimilation patterns of these speech contrasts, if any, before and after training in individual language groups as well as across groups. Of these, the third goal is descriptive and is di scussed in the last section of this chapter. The design of the experiment, mainly the choice of the subject groups, has its basis on the hypothesis formed for this study. It is expected that a positive influence of m ulti lingual benefit on the acquisition of speech contrast will be seen in both bilingual groups through their better performance in Identification and discrimination tasks when compared to monolingual group. The factor of feature generalization is hypothesized to show an additive effect over m ulti lingual benefit in the acquisition of the non native contrasts presented. Also, a shift in perceptual assimilation types reflecting a trend towards learning is hypothesized in relation to the third goal of the study. The appropriateness of the choice of la nguage groups in order to examine the first two goals has been discussed in previous chapters. The a priori possibilities from the three as results of this study de pending upon different hypotheses stated below. P ossibility (1) results if the presence of multilingual benefit emerges as dominant through the perceptual performance of both bilingual groups, that is, Bengali English (BE) and Spanish English (SE), will be significantly higher than that of the monolingual group, American English subjects (AE). P ossibility (2) reflects the scenario where only BE group may show higher

PAGE 71

71 perceptual performance than either the SE or AE group, evidencing that only feature generali zation is facilitating the BE group to learn to perceive the non native speech contrast faster than the rest. There will be no evidence for multilingual benefit effect if the SE group fares as equal to AE, the mo nolingual group in this case. Possibility (3 ) is concerned with the assumption that feature generalization may be seen as an additive effect over the multilingual benefit factor. In this case then, the BE group will perform significantly higher than the SE group which in turn will perform higher tha n the AE group. In other words, the BE group will have added advantage of enhancing the learning process since both multilingual benefit as well as feature generalization factors will facilitate this process. P ossibility (4) represents the fourth logical a lternative, that of null hypothesis. In this case, there will not be any significant changes in the perceptual performances of all the three groups displaying no evidence for either multilingual benefit or feature generalization. Three different tests have been employed to investigate the hypotheses posited above. Data of a total of sixty subjects, twenty from each language group, have been analyzed. The perceptual performance of subjects in various tests like the identification test, AX discrimination test and perceptual assimilation test ha s been measured by comparing the performance at the pretest level to that at the posttest level. In order to normalize any performance differences at the pretest level, difference in the scores of the posttest from the p retest have been submitted as data to assess significance level. The results presented in this chapter were tested for statistical significance using a number of mixed model ANOVA (Analysis of Variance) on various tests employed during pretest and posttest phase of the experiment. The independent variables for the AX discrimination test, identification tests (Pre post ID task and Generalization task) as well as the perceptual

PAGE 72

72 assimilation test remained same, such as native language group and contrast type. ANOVAs reported the main effects of the independent variable, language group, as well as the interaction between these variables, language group x contrast. In the discrimination test ANOVAs, the dependent variable was mean percent correct score of a subje ct on sixteen trials of a particular contrast type. In the varied identification test ANOVAs, the dependent variable was mean percent correct score of a subject on twelve trials of a particular contrast. The percent change in category assimilation type was used as the dependent variable for the perceptual assimilation marginally significant such as p<0.1, then Tukey post hoc tests were conducted t o examine the differences of the least squares means of the fixed effects in order to determine the statistical significance of performance by individual language groups. Results This section gives a detailed account of the results in the areas of testing, such as Iden tification task, AX discrimination task, generalization test and the perceptual assimilation task. Since the results of perceptual assimilation require a more descriptive approach than a statistical analysis, they are discussed in a separate sub section be low. Once the overall results containing the main effect of independent variables are stated for the test of identification, discrimination and generalization (provided immediately below), specific post hoc results are discussed in detail under the heading s for the relevant factors examined in this study. Main Effects and Interaction of Independent Variables In the case of the identification test on a trained on talker, comparing the results of pretest and posttest, no main effect for language groups was ob served ( F(2,57)=2.42, p>0.05). Fig 3 1 shows the mean percent correct response in pretest and posttest level of identification test by subjects from the three language groups. All three listener groups improved from pretest to

PAGE 73

73 posttest, with the largest i ncreases in performance observed with the Bengali English group, followed by the Spanish English and American English trainees. The main effect for contrast type was observed to be significant, as expected ( F (3,171)=5.46, p<0.05). The interaction between l anguage groups and contrast was observed to be statistically significant (F (6,171)=5.74, p<0.05). Figure 3 1. Mean percent correct identification scores averaged over all four contrasts at the pretest and the posttest level. Since the language group contrast interaction was significant, post hoc tests using were conducted for only lateral nasal and fricative contrast types. The rhotics contrast showed no group differences. The se results of individual contrasts are discussed in p ost hoc analysis under the section of individual factors. The results of perceptual performance from the comparison of pretest and posttest level of AX discrimination test show no main effect for language group ( F(2,57)=1.55, p>0.05, n.s). Figure 3 2 repr 51% 61% 56% 63% 80% 74% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% AE BE SE % Correct Response Language Groups AE=American English monolinguals; BE=Bengali English bilinguals; SE=Spanish English bilinguals Pretest Posttest

PAGE 74

74 level for the three language groups. All three listener groups improved from pretest to posttest, with the largest increases in performance observed with the Bengali English gr oup, followed by the Spanish English and American English trainees. The interaction between language group and contrast was observed to be significant (F(6,171)=4.04, p<0.05). Of the post hoc analysis using Tukey test, only the lateral contrast showed stat istical significance for language group effect. Figure 3 2. In the case of the generalization test (an identification t ask), the scores were compared to scores of the identification test (posttest) test in order to examine any evidence for the development towards forming new phonetic categories for the non native sounds by subjects in any of the language groups. Posttest s trained on talker while the generalization scores displayed the accuracy in identifying the stimuli from the novel talker (a voice not heard before during the experiment). The differenc e of the posttest scores and generalization scores were submitted as data for the ANOVA test. The results for the test of generalization, where similar stimuli from a new talker were tested for accurate identification, showed no main effect for language gr oups ( F(2,57)=1.10, p>0.05, n.s). The 0.85 1.17 0.95 1.30 1.94 1.64 0.00 0.50 1.00 1.50 2.00 2.50 3.00 AE BE SE Subjects' Averaged D' Scores Language Groups Pretest Posttest

PAGE 75

75 interaction between language group and contrast was observed to be statistically insignificant (F(6,171)=1.21, p>0.05, n.s). None of the contrasts showed any significance concerning the language groups. Therefore, no post hoc testing was conducted for the test of generalization. Mean identification scores of the generalization test when compared with the posttest scores, displayed unexpected results. Previous studies have provided evidence of development of robust pho netic categories by L2 learners through generalization tests (Lively et al. 1994). Based on their results, it was expected that subjects will perform either equally well or worse in identifying the stimuli from a novel voice than the trained on voice, as w as the case in posttest ID test. If any effects of multilingual benefit and/or feature generalization were to be observed, then the expected results would be that either the BE (for feature generalization) or both the BE and the SE groups (for multilingual benefit ) would perform equally well providing evidence for the development of separate phonetic categories for the non native contrasts. On the other hand, the AE group, acting as the control group for both factors, would perform significantly worse in t he generalization test displaying that AE group may require more training or exposure to these contrasts in order to develop separate phonetic categories. The results from the current experiment in terms of mean identification scores for the generalizatio n test are provided in Figure 3 3. For all three listener groups, the identification of the non native contrasts was more accurate in tokens from the new talker when compared with the trained on talker, with the slightly larger increase in performance obse rved with the Bengali English group, followed by the Spanish English and American English trainees.

PAGE 76

76 Figure 3 3. M ean percent correct identification of contrasts spoken by trained on t alker (posttest) and new talker (generalization test). Therefore look ing at the overall results of the tests used to factor out any presence of facilitating effects of multilingual benefit and/or feature generalization, it is difficult to conclude that any of the hypotheses have been proven true. However, some of the post h oc analyses of the ANOVA results, where marginal significance is observed in either main effect for language groups or interaction between language groups and contrasts, are discussed in the following section. These analyses may shed some light on the not so robust (perhaps subtle) effects of the factors explored in this study. Examining Evidence for Individual Factors Multilingual benefit As we have seen in the above section, the main effect for language in the pre post identification test was on the borde rline of significance level ( F (2,57)=2.42, p>0.05) (p=0.09). The averaged scores over all four contrasts at the pretest and posttest level reflect this trend. Looking at figure 3 4, we can see that both bilingual groups show a noticeable increase in the 63% 80% 74% 64% 82% 75% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% AE BE SE % Correct Response Language Groups Trained on Talker Novel Talker

PAGE 77

77 a acting as the control group. Figure 3 4. Individual comparison of the monolingual group with BE bilingual group (top chart) and SE bilingual group (bottom c hart). The mean percentage of improvement in the bilingual BE group is calculated as 19% whereas the improvement percentage in monolingual AE group is observed as 12%. The mean percent improvement in accuracy scores of the bilingual SE group is also seen a s 19%, identical to that of the bilingual BE group. 51% 61% 63% 80% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% AE BE % Correct Response Language Groups Pretest Posttest 51% 56% 63% 74% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% AE SE % Correct Response Language Groups Pretest Posttest

PAGE 78

78 These differences of the posttest result from that of pretest results reflect the actual perceptual performance by normalizing any initial performance differences at the pretest level. Looking at the aver aged improvement for all three groups, the prediction that both bilingual groups individually have higher performance level than monolingual group appears to be acceptable. In addition to the above results, the interaction between language groups and contr asts was significant ( F (6,171)=5.74, p<0.05) (p=0.0001). Therefore, post hoc analysis was conducted using Tukey procedure. Apart from the rhotic non native contrast, all other contrasts (laterals, nasals, fricatives) showed significance for the language g roup*contrast interaction. The lateral non native contrast was observed to be significan t ( F (2,171) =7.62, p<0.05). Similar significance results were observed for the nasal non native contrast (F (2,171)=4.54, p<0.05) and the fricative non native contrast (F (2,171)=3.75, p<0.05). Of these, the only lateral non native contrast was observed to show any evidence for the facilitative effect of multilingual benefit in bilingual groups (BE and SE) since the performance level of identification of the lateral con trast was significantly more among bilinguals than the monolingual group The nasal non native contrast reflects a trend towards feature generalization. Therefore the nasal along with the fricative contrast is discussed in detail under the next section. N ext, in order to determine the significant differences between each of the language group pairs, the differences of least squares means were obtained using the post hoc Tukey test. When analyzing the post hoc results for the lateral non native contrast, it was observed that the perceptual performance of accurately identifying the sounds in this non native contrast by the Bengali English bilinguals was significantly better than the American English monolinguals (p<0.01). The Spanish accurate perception of the lateral contrast was

PAGE 79

79 marginally different than that of the American English language group (p = 0.09). Figure 3 5 shows the mean accuracy scores of identification of the lateral contrast at the pretest and posttest levels by the three language groups. The mean percent improvement in performance by BE group was the largest (25%), followed by SE group (18%) and AE group (8%). On the other hand, when the differences of least squares means of the bilingual groups, BE and SE, were com pared using the Tukey post hoc test, the pair showed no statistical significance in their performance from pretest to posttest level (p = n.s). Figure 3 5. Mean percent correct response in identifying the lateral non native contrast [ ] in the iden tification test at the pretest and posttest level. These results for the lateral non native contrast provide evidence for the facilitative effect of multilingual benefit since Spanish speaking subjects, without any linguistic experience with retroflex feat ure, were able to perform nearly as well as Bengali speaking subjects. The monolingual language group (AE) was significantly (marginal with SE group) different from both the bilingual groups. 47% 65% 56% 55% 90% 74% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% AE BE SE % Correct Response Language Groups Pretest Posttest

PAGE 80

80 Few of the post hoc testing for the AX discrimination test is s uggestive of multilingual benefit factor. As stated before, the main effect for language group in the improvement of performance in case of discrimination test was not significant. However, the interaction between language groups and contrasts was observed to be statistically significant. Only the lateral non native contrast was observed to have significant interaction (F (2,171)=9.86, p<0.05). The non native contrasts with manners as nasal (F (2,171)=0.46, p= n.s ); fricative (F (2,171)=0.62 p= n.s ); as wel l as the rhotics (alveolar tap v er s us retroflex approximant) non native contrast (F (2,171)= 0.04, p= n.s ) were observed to have no significant interaction of language groups with contrasts. Analysis of the differences of least squares means in the post hoc Tuke y test for the lateral non native contrast, showed that the improvement in performance by both bilingual groups (BE and SE) individually was statistically significant over the performance by the monolingual group (AE). Improvement in the Bengali speaking s native contrast was more than the monolingual American English speakers (p < 0.00 1 ). The Spanish speakers discriminated the lateral non native contrast better than the monolingual American English speakers (p < 0. 0 5 ). However, the discriminability of the lateral contrast when compared between the bilingual groups showed no significant difference (p = n.s ). These results show that the bilinguals were performing significantly better than the monolinguals in the case of lateral contrast perhaps implying the presence of multilingual benefit effect. Figure 3 6 represents the groups.

PAGE 81

81 Figure 3 6. ion of the lateral non native contrast [ ] at the pretest and posttest levels by the subjects in the three language groups. Looking at the post hoc analysis of the lateral non native contrast in the identification test, which was the same task as used d uring training, as well as the AX discrimination task provided evidence for a trend that may direct towards demonstrating that among bilinguals, multilingual benefit may perhaps be at work in enhancing the acquisition of the non native contrasts more than is observed in monolinguals. Feature g eneralization Feature generalization/feature productivity refers to the previous linguistic experience with a native phonetic feature that may aid in the acquisition of a non native contrast that exploits this same fea ture but differs from any native contrast along other dimensions. One of the hypothesis postulated as outcome for this study was that the Bengali English group may perform significantly better than the Spanish English group due to its experience with a ret roflex feature in native contrasts. Additionally, both bilingual groups may perform significantly better than the monolingual American English group. In this case, feature generalization is expected to have an additive effect over multilingual benefit fact or in this case. 1.00 1.58 0.86 1.04 2.77 1.75 0.00 0.50 1.00 1.50 2.00 2.50 3.00 AE BE SE Subjects' Averaged D' Scores Language Groups Pretest Posttest

PAGE 82

82 It implies that when compared, within the bilingual groups, the Bengali English group is expected to perform significantly better than the Spanish English group. However, the results showed that feature generalization was not a significant factor. In most of the ANOVAs and the Tukey post hoc analysis, no significant differences were observed In the case of Tukey post hoc analysis of the nasal non native contrast duri ng the identification test at the pretest and the posttest level t significantly more than the American English speakers (p < 0.0 5 ) but not significant from those of the Spanish speaking subjects (p = n.s ). The statistical significance was also not observed for the Spanish En glish bilinguals and the American English monolinguals (p = n.s ). Figure 3 7. Mean percent identification scores for the nasal non native contrast [ ] by the subject gr oups at the pretest and the posttest level. Figure 3 7 shows the averaged identification scores for the nasal non native contrast by the language groups at the pretest and posttest level. All three listener groups improved from pretest to posttest, with th e largest increases in performance observed with the Bengali English group, 51% 59% 54% 59% 80% 68% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% AE BE SE % Correct Response Language Groups Pretest Posttest

PAGE 83

83 followed by the Spanish English and American English trainees. These results may suggest the possibility that perhaps feature generalization may have been at work facilitating the Bengali speakers to generalize the retroflex feature to the nasal non native contrast. One area where statistical significance is seen between Bengali English and the Spanish English language group (p < 0.0 5 ) is the post hoc analysis of the fricative (pala tal vs. retroflex) non native contrast within the pretest posttest identification test ANOVA. Figure 3 8 provides the averaged scores of identification of the fricative non native contrast during the pretest and posttest levels. All three listener groups i mproved from pretest to posttest, with the largest increases in performance observed with the Spanish English group, followed by the Bengali English and American English trainees. Figure 3 8. Mean percent identification scores for the fricative non nati ve contrast [ ] by the three language groups at the pretest and the posttest level. this particular contrast is similar to that of the monolingual group (p = n.s). This result does not support the feature generalization hypothesis since the statistical significance between the BE 47% 52% 46% 57% 62% 67% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% AE Be SE % Correct Response Language Groups Pretest Posttest

PAGE 84

84 and the SE groups is showing inverse results (SE>BE). This result does not support the multilingual benefit hypothesis also since BE biling uals did not perform any better than the AE monolinguals. The interpretation of these results is discussed in chapter four. The se results are suggestive of the effect of multilingual benefit but not of the effect of feature generalization factor proposed in the hypothesis. The feature generalization effect, expected to be seen in BE bilinguals because of their linguistic experience with retroflex feature, appears to have no additive effect over the general factor of multilingual benefit expected in bot h the bilingual groups. Therefore, any significant differences seen between the BE and the AE language groups can be suggestive of multilingual benefit more so than of feature generalization. The effect of multilingual benefit factor in enhancing the acqui sition of non native contrast is observed exclusively in the non native contrast with lateralization as the manner. The lateral non native contrast was considered the most difficult contrast for this study since all the three language groups have only a si ngle phonetic category for a lateral sound. Therefore, perhaps the subtle presence of multilingual benefit and its effects seen in the surfaced in the acquisition of relatively difficult contrasts. The results seen with latera l non native contrast may reflect the positive influence that is not prominent enough to be observed with easily discriminable contrasts. Perhaps by increasing the sample size of the language groups may bring forth the factor of multilingual benefit a s a robust effect. Therefore, to inspect for more robust effects regarding the multilingual benefit factor, a reanalysis of the experimental data was conducted. The reanalysis is discussed in the following section. Re Analysis As seen in the results above, none of the tests or individual contrasts reflect any robust effects of feature generalization. However, the presence of multilingual benefit is marginally

PAGE 85

85 evident in the perceptual performance (both identification and discrimination tests) of the lateral non native contrast and as an overall result suggesting that the scenario of the bilingual groups (BE and SE) performing better at identifying the non native contrasts than the monolingual group (AE) is accurate. These results, although not robust enough to be conclusive, provide us with reasonable doubt and reflect towards a trend of the subtle presence of multilingual benefit playing a role in enhancing the acquisition of non native contrasts, particularly the difficult place distinctions. Therefore, it is appropriate to further explore this area in a different light. Since the results indicated improved perceptual performance by bilingual groups over the monolingual group, it was thought suitable to examine the factor of bi/multilingualism as a whole ra ther than two separate individual groups. Therefore, a re analysis of the present data (from the two tests provided at pretest and posttest levels: Pre Post ID test and AX discrimination test) was conducted. Re analysis: Multilingual Factor In this re anal ysis, data from both the bilingual groups, Bengali English and Spanish English were put collectively under the bilingual group (N=40). The American English subject groups was reiterated as the monolingual group (N=20). The pooled data was submitted to a mi xed model ANOVA test where the independent variable, this time, was bilingualism vs monolingualism, instead of individual language groups. The results of all the tests mentioned above were re examined with this ANOVA design. Figure 3 9 represents the aver aged accuracy scores for identifying all four non native contrasts at the pretest and the posttest level by the bilingual groups (BE and SE) and the monolingual group (AE). The improved performance in percentage seen in subject groups after training was 19 % for the bilingual group and 12% for the monolingual group across all four contrasts.

PAGE 86

86 Figure 3 9. Mean percent scores of identification test across all four contrasts by the bilingual groups (BE and SE) and the monolingual group (AE) at the pretest and the posttest levels. effect for multilingualism showed statistic al significance (F (1,58)=4.75, p <0.05). The interaction between the multilingualism and contrast was also significant (F (3,174)=4.45, p<0.05). On examining the tests of effect slices, it was observed that of the four contrasts, the lateral non native co ntrast (F (1,174)=11.72, p<0.05) and the nasal non native contrast (F (1,174)=5.03, p<0.05) showed statistical significance between the independent variable, bilingualism monolingualism. In case of the AX discrimination test results, the pooled data set of revealed marginal ly different from that of monolingual subjects. collectively compared to those of monolingual subjects has been represented in figure 3 10. The 51% 58% 63% 77% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% Mono Biling % Correct Response Language Groups Pretest Posttest

PAGE 87

87 main effect for multilingualism was borderline (F (1,58)=2.95, p = 0.09) unlike the insignificant results (p = n.s ) of discrimination test when examining individual language group effects Figu re 3 groups (BE and SE) and the monolingual group (AE) at the pretest and the posttest levels The interaction between group and contrasts was observed to be significa nt (F(3,174)=6.59, p<0.05). Therefore, post hoc analysis was conducted on individual contrasts using the Tukey procedure Of the four contrasts, only lateral non native contrast showed statistical significance between the bilingual and the monolingual grou p s (F (1,174)=18.50, p<0.05). The post hoc results from the re analysis were quite similar to the post hoc results from the main analysis in terms of which contrasts showed significant differences among the independent variables The results from the poole d dataset exhibit the effects of multilingual benefit being employed by the bilinguals in order to enhance the acquisition of the non native contrasts. However, this re analysis reveals the relevance of certain methodological issues. The effects of certain factors like multilingual benefit perhaps cannot be captured at the standard sample size of 0.85 1.06 1.30 1.79 0.00 0.50 1.00 1.50 2.00 2.50 3.00 Monolingual Bilingual Subjects' Averaged D' Scores Language factor Pretest Posttest

PAGE 88

88 twenty subjects used in this study. These effects could be subtle and perhaps obscured by individual subject variations. This limitation is further discussed in ch apter four. An A lternative R e analysis: Only Place Contrasts Alternative re analysis of the data was conducted by examining individual contrasts used for this study. The results from the main analysis and subsequently the results from the pooling analysis reflected the need to investigate individual contrasts and to look for inconsistencies among the four contrasts chosen for this study. The results that looked at individual language groups showed enormous amount of variability in the means procedure of AN OVAs conducted. In addition to that, the perceptual performances varied from one contrast to the other. Based on the three reasons stated below, a re examination was conducted of the four non native contrasts as hindsight. Looking at the preliminary experi ment, where data of initial AX discrimination of these contrasts were gathered from native American English speakers, t attained by the rhotics non 2 pg 18, for details). Though not at cei discriminability at above chance (63%) whereas the other three contrasts showed discriminability at chance level (50%) Therefore, ease of discriminibility of this contrast may have s uppressed since the statistical analyses were conducted on the difference scores averaged over all contrasts This possibility gets support from the pretes t scores for the rhotics contrast which were well above chance level for all three groups (AE=59%, BE=67% and SE=67%). Thus, the probability to observe subtle effects of multilingual benefit and/or feature generalization would have been minimal.

PAGE 89

89 Lookin g more closely at the phonetic features of this particular contrast, it appears to evade the overall homogeneity of the stimuli for this study. The non native contrasts with manners lateral approximant, nasal and fricative only differ in terms of place dis tinctions. The lateral non native contrast has an alveolar and a retroflex [ ]. The nasal non native contrast has an alveolar and a retroflex [ ] (see chapter Two for detailed discussion on dental/alveolar distributions in Malayalam). The fricative non native contrast has a palato alveolar and a retroflex [ ] distinction. However, the rhotic non native contrast [ ] differs in not only the place distinction but also the manner. This contrast involves an alveolar tap and a retroflex approximant. There fore, the rhotic contrast stands apart from the rest of the three only place contrasts. Perhaps this could have provided the contrast with inherent extra salience of acoustic features which may have provided the subjects with ease of discriminability. This interpretation might raise questions on the reasons for choosing rhotics contrast as one of the stimuli. In hindsight, initially it was assumed that feature generalization might be a robust factor and will thus emerge even if there were additional seconda ry cues to the contrast involving manner or other features that may provide easier discrimination of the contrast. However, since this is not the case as is evident from the main results, a reanalysis exclusively of the place contrasts is called for. One m ore point of contention within this particular contrast is that the phonetic definition of retroflex approximant sound in Malayalam has been a debatable issue among the phoneticians researching this language (see chapter Two, pg 6, for details). In hindsig ht, no particular phonetic description has been agreed upon making this particular contrast of concern regarding as stimuli for this study. Therefore, a second re analysis of the data set where only three non native contrasts were analyzed was conducted. In this case, the scores on the non native contrasts which differed only

PAGE 90

90 in place distinctions (lateral, nasal fricative ) were used in a mixed model ANOVAs for three datasets: Identification test, AX discrimination test and the generalization test. In the case of the identification tests at the pretest and posttest levels, the main effect for language groups was observed to be significant (F (2,57)=5.12, p<0.05). Figure 3 11 shows the averaged accuracy scores of identification test over the three place contrasts by all language groups at the pretest and posttest levels. All three listener groups improved from pretest to posttest, with the largest increases in performance observed with the Bengali English group, followed by the Spanish English and America n English trainees. The improvement percentage for the bilingual groups was significantly higher. For the BE group it was 19% and for the SE group it was 17%. The percentage of improvement in perceptual performance seen in AE group was 9%. Figure 3 11. Mean percent correct response for identification of the three place contrast (laterals, nasals, fricatives) by bilingual (BE and SE) and monolingual (AE) groups at the pretest and posttest levels. The interaction of language groups with the three contrasts was also observed to be significant (F (4,114)=5.81, p<0.05). The main effect of contrast showed no significant 48% 59% 52% 57% 77% 69% 0% 10% 20% 30% 40% 50% 60% 70% 80% 90% 100% AE BE SE % Correct Response Language Groups Pretest Posttest

PAGE 91

91 difference, unlike the main analysis results, providing evidence for homogeneity within the three place contrasts (F (2,114)=1.81, p= n.s.) (p= 0.16). The post hoc analysis using Tukey test was conducted. All three non native contrasts showed significant differences between the language groups. The post hoc results were the same as the results from the main analysis since there was no change in th e data set within the individual contrasts. In the case of AX discrimination test, the AN OVA results indicated marginal difference between language groups. Figure 3 three non native contrasts a t the pretest and posttest levels. All three listener groups improved from pretest to posttest, with the larger increases in performance observed with the Bengali English group, followed by the Spanish English and American English trainees. The main effect for language groups was seen at the borderline significance (F (2,57)=2.81, p=0.06). The interaction between the language groups and the three place contrasts was observed to be significant (F (4,114)=5.45, p<0.05). Figure 3 12. e AX discrimination tests for the three place contrasts (Laterals, nasals, fricatives) by the language groups at the pretest and posttest levels. 0.70 1.15 0.67 0.97 1.87 1.18 0.00 0.50 1.00 1.50 2.00 2.50 3.00 AE BE SE Subjects' Averaged D' Scores Language Groups Pretest Posttest

PAGE 92

92 The results for the generalization test scores when compared with the posttest scores showed no statistical si gnificance, similar to the main analysis results. The area of testing for robust phonetic category development did not reveal any significant differences among the language groups in any of the analyses. One explanation could be the within group variabilit y. Another explanation could be the issue of baseline talker intelligibility. These various explanations will be discussed in detail in chapter four. In summary, co nducting re analyses of the data demonstrated a different pattern of results when 1) both bilingual groups were pooled as one group and 2) problematic contrast was excluded. The positive influence of the multilingual benefit factor was supported with the d ependent measures when the issue of sample size was accounted for by pooling the BE and SE groups together. The results for the identification tests showed significant group differences between bilinguals and monolinguals and those of AX discrimination dis played marginal significance in the performance of bilinguals over monolinguals. The reanalysis with only place contrasts also revealed a significant main effect for language groups. However, effects of the feature generalization factor did not surface pro minently even in the reanalysis results. The next section presents the results of the perceptual assimilation tests conducted at the pretest and posttest levels during the training experiment. Perceptual Assimilation Results The perceptual assimilation ta sk conducted at the pretest and posttest levels requires a more descriptive interpretation of its results. Therefore, this section presents results solely from these tests. Because of its descriptive nature, no prior predictions could be made on the resul ts of perceptual assimilation patterns. However, based on PAM L2 predictions (Best & Tyler 2007), we could expect to see a direction towards learn ing For this study, learn ing was defined in terms

PAGE 93

93 of a shift from non categorical assimilation types like single category (SC) or UU (both uncategorizable phonetic sounds) to assimilation types such as category goodness (CG) or uncategorizable and categorizable type (UC) which reflect a trend towards higher sensitivity of discrimination or even up to two category assimilation which represents excellent discrimination of the non native contrasts. The following sections describe results from the perceptual assimilation test in terms of changes in assimilation types after training across all language groups. In addition, the question of whether the effects of factors like multilingual benefit are seen in terms of perceptual through the difference scores of assimilation types elicited from th e pretest and posttest results. Changes in Assimilation Types after Training For this study, the assimilation types (for details, refer to chapter 1 ) described in PAM model have been applied for interpreting the results. Five major as similation types which reflect various levels of assimilation that is, two category (TC), single category (SC), category goodness (CG), uncategorizable categorizable (UC) and both uncategorizable (UU) have been used. The results from the perceptual assimi lation test reveal a consistent shift of assimilation types from none / minimal separate categorization of non native contrasts before training to assimilation types like TC, CG and UC which reveal improvement taking place in discrimination of the non nat ive contrasts during training This implies a direction towards learning of the non native contrasts as separate phonetic categories. Figure 3 13 shows the mean percentage of number of assimilation types drawn from subjects over all three language groups c ollectively. The results showed a decrease of 18% in SC assimilation type and 5% in UU assimilation type from pre training to post training levels. The TC and CG assimilation types displayed an increase of 11% and 10% respectively. The UC

PAGE 94

94 assimilation type showed a modest increase of 2% from pretest to posttest. The overall change of assimilation types from showing less categorization to a higher sensitivity towards categorization provides evidence of learning taking place during the limited training period The changes in these assimilation types were not similar for all the four contrasts. Figure 3 13. Mean percentage of assimilation elicited from all three language groups, pretest vs posttest. Looking at the results for the four non native contrasts (f igure 3 14) across all language groups, decrease in SC assimilation type was observed for all contrasts. For the lateral and the nasal non native contrasts, there was found to be a steep decrease ( 23% and 29% respectively) when compared to the rhotic or the fricative contrasts ( 12% and 10% respectively). This steep increase is manifested as a large increase in CG (22% and 17% respectively) and TC (11% and 14% respectively) assimilation types for the nasal and lateral contrasts. This provides evidence of possible learning of the non native contrasts taking place. The rhotic contrast showed a modest increase in CG (2%) but a comparable increase in TC (13%) assimilation types. In case of UU (2%), it showed an increase unlike other non native contrasts. The fricative contrast, on the other hand, showed very modest increase in CG (2%) and TC (7%) types which were expected to 46% 17% 12% 16% 9% 28% 27% 7% 18% 20% 0% 5% 10% 15% 20% 25% 30% 35% 40% 45% 50% SC CG UU UC TC Perceptual Assimilation Types Pretest Posttest

PAGE 95

95 show learn a b i l i t y Nevertheless, it displayed learn a b i l i t y through UC assimilation type where the positive difference (improvement) is 12 %. Figure 3 14. Percentage of difference scores (pretest to posttest) of assimilation types for the individual contrasts elicited from all three language groups. On the whole, the SC and UU assimilation types displayed a decrease from pretest to posttes t whereas the CG and TC types showed an interesting increase. Overall, the subjects showed more learning in lateral and nasal non native contrasts than the other two contrasts. Changes in As similation Types: Multilingual Benefit F actor The p erceptual assim ilation test was one of the areas to examine the effects of factors like multilingual benefit and feature generalization on the acquisition of non native contrasts. These chosen contrasts were phonemically assumed to be assimilated as the single category t ype at the initial exposure (pretest) level. The assimilation results from the previous section provide evidence of a definite shift towards learning these novel non native contrasts in a very limited training period. However, the question arises whether a ny language group differences were seen pointing towards the effects of factors explored in this study. 11% 4% 6% 22% 23% 14% 4% 5% 17% 29% 13% 5% 2% 2% 12% 7% 12% 10% 2% 10% 30% 20% 10% 0% 10% 20% 30% TC UC UU CG SC Perceptual Assimilation Type Lateral Nasal Rhotic Fricative

PAGE 96

96 The results from the identification tests and AX discrimination tests revealed no evidence supporting the presence of feature generalization as a facili tating factor in the acquisition process by the bilingual group, BE. Marginal effects were observed for multilingual benefit in case of the lateral non native contrast. On two alternate reanalysis of the data set, the presence of this factor emerged as sig nificant for the bilingual group when compared with monolingual group. Therefore, it is reasonable to examine the bilingual groups as a whole comparable to monolingual group rather than individual language groups. Figure 3 15 displays percentage of assimi lations elicited from individual language groups at the pretest and posttest levels. The bilingual groups, BE and SE, showed similar shifts in assimilation types from pretest to posttest, unlike the monolingual group. The SC type of assimilation pattern wa s observed to have similar percentage at the pretest level for both BE (38%) and SE (41%) groups, whereas the same assimilation type had a very high percentage of 60% at the pretest level for the monolingual AE group. Interestingly, a sharp decline was see n in the SC type at the posttest level for both bilingual groups, unlike the monolingual group. An inverse trend was observed for the TC and in some cases for the CG assimilation type from pretest to posttest. Both bilingual groups showed a high rise in th e TC assimilation type at the posttest level contrary to the monolingual group. The SE group was deviant from the BE group in case of UC assimilation type. The SE group displayed a modest decrease of ( )5% whereas the other two groups showed similar percen tage but increase in UC type. Taking a comparable view of the three groups, figure 3 16 shows the difference scores of the percentage of perceptual assimilation types comparing the two bilingual groups with the monolingual group over all four contrasts. A positive percentage denotes an increase in the

PAGE 97

97 Figure 3 15. Percentages of assimilation types elicited from individual language groups at the pretest and posttest levels are provided in three different charts. 10% 14% 4% 13% 60% 14% 18% 0% 23% 46% 0% 10% 20% 30% 40% 50% 60% 70% TC UC UU CG SC Perceptual Assimilation Types American English Monolinguals Pretest Posttest 6% 14% 16% 26% 38% 23% 20% 10% 34% 14% 10% 0% 10% 20% 30% 40% 50% 60% 70% TC UC UU CG SC Perceptual Assimilation Types Bengali English Bilinguals Pretest Posttest 11% 20% 16% 11% 41% 25% 15% 11% 25% 24% 10% 0% 10% 20% 30% 40% 50% 60% 70% TC UC UU CG SC Perceptual Assimilation Types Spanish English Bilinguals Pretest Posttest

PAGE 98

98 percentage of elicited assimilation ty pe from pretest to posttest and a negative percentage denotes a decrease in the assimilation types gathered by the language groups. Figure 3 16. Percentage of difference scores (posttest pretest) of assimilation types elicited from BE, SE bilingual grou ps and AE monolingual group over all four contrasts. A steep decline was observed in the SC type along with a nearly equal steep increase in the TC assimilation type, and CG type for SE group, during the posttest for bilinguals. On the other hand, the mono lingual group showed a decline in SC type lower in percentage than that of the bilinguals. Moreover, the SC type remained high in percentage even at the posttest level for the monolingual group. Only a modest increase was seen in the TC, UC and CG assimila tion types which reflected limited learning taking place within that group. Thus, the results described from the perceptual assimilation tests provide evidence for the effects of multilingual benefit facilitating the acquisition of novel non native contras ts even with such limited training. Although this section seem ed to have required a more descriptive analysis of the perceptual assimilation results, unlike the statistical results from other tests such as the 4% 4% 4% 10% 14% 16% 6% 6% 8% 24% 14% 5% 5% 14% 18% 25% 20% 15% 10% 5% 0% 5% 10% 15% 20% 25% TC UC UU CG SC % difference scores (Posttest Pretest) Perceptual Assimilation Types Ae Be Se

PAGE 99

99 i dentification test and the discrimination tes t, a statistical analysis (one way ANOVA) on proportion change in assimilation types was conducted to examine significance based on the individual language groups as well as the multilingual factor. The main effect for language was observed to be significa nt (F (2,57)=4.16, p<0.05). In the post hoc T ukey test, BE group showed proportionately more change in assimilation types than the AE group (p <0.05 ). The BE and SE group showed no significant difference in proportion change (p = n.s ). The SE group showed no significan t difference from AE group for proportion change (p = n.s ). These result s lead to the reanalysis of the dataset by pooling the bilingual groups together and comparing with the monolingual group. To look at the multilingual factor, a one way AN OVA was conducted which showed significant differences between the bilingual group and the monolingual group (F (1,58)=7.78, p<0.05) These results, again, are suggestive of the effect of multilingual benefit on acquiring the non native contrasts by the bi linguals as a group, regardless of their L1 languages. Summary The results of the current study are suggestive of the effects of multilingual benefit The main language effects in reanalysis and post hoc tests of the lateral non native contrast provided s upport for the hypothesis stated for the multilingual benefit factor. The evidence for this factor was mainly seen in the lateral non native contrast in the dependent measures like the identification test and the AX discrimination test. Both t he reanalysis provided evidence for multilingual benefit factor by showing significant group differences. However, neither the effect of feature generalization nor its additive effect surfaced in any of the results. The only marginal effect of feature generalization wa s seen in the identification test results for the nasal non native contrast which does not show strong support for this factor. The test of generalization results were observed as non significant in terms of group differences as the mean percentage of

PAGE 100

100 per formance revealed that on an average subjects performed better with the novel voice (generalization test) than with trained on voice (posttest). The results from the perceptual assimilation tests revealed a consistent shift towards learning of the non nati ve contrasts. Also, they showed trends of effects of multilingual benefit among bilingual groups as opposed to the monolingual group. These results of this very limited training are only suggestive, of the hypothesis that bilinguals can acquire non native contrasts faster than m onolinguals given the facilitating effects of multilingual benefit However, no study is without its caveats. There were some unexpected results such as subjects did better with novel voice than a trained on voice. In addition, a lo t of within group variability was found for the generalization test. All the results, their interpretation and caveats are discussed in the next chapter.

PAGE 101

101 CHAPTER 4 DISCUSSION AND CONCLUSION Introduction The experiment was designed to address a number of im portant questions concerning third language acquisition and cross language speech perception: Would multilingual benefit and/or feature generalization facilitate the acquisition of novel non native contrasts among bilinguals? Can perceptual assimilation pa tterns be modified through a limited training period? Would these changes in assimilation patterns, if any, reflect differences between the bilingual and monolingual subjects in terms of acquisition of non native contrasts? These broad questions are addres sed through an overview of the results. In addition to that, a discussion regarding the possible explanations for the unexpected results seen in this study will be put forth discussing some unforeseen caveats in the methodology as well as issues with curre nt assimilation models. The results of this study will also be discussed in tandem with previous literature in order to embed the results in a larger perspective, to further future potential hypotheses in the field of language acquisition. Overview of the Results The design of the current study had three main aims. The f irst aim concerned the assessment of the effects of metalinguistic ability developed as a result of having two languages already in place T he second aim pertain ed to the assessment of the effects of previous linguistic experience with certain phonetic features present in the target stimuli. The third aim was to describe changes in assimilation patterns and examine any trends towards learnability. These aims were narrowed down to more specif ic hypotheses. The choice of subject groups was determined based on the formulation of the hypotheses. Briefly stated, Bengali English (BE) bilinguals were chosen because of their linguistic experience with the retroflex phonetic feature

PAGE 102

102 and more so becaus e of the fact that they were balanced bilinguals. Spanish English (SE) bilinguals were appropriate as the counteracting group solely because of their lack of any linguistic experience with retroflex feature and also for having acquired two languages relati vely simultaneous. American English (AE) language speakers were chosen as the control group due to lack of any substantial knowledge of another language, making them a monolingual group. Based on these language groups, more strictly defined hypotheses conc erning the main aims were constructed. The effects of multilingual benefit would be evident if both bilingual groups perform ed significantly better at identifying the non native contrasts than the monolingual group. The facilitating effects of feature gene ralization would be seen if the BE group performed significantly better than the SE and the AE group. Additive effects of feature generalization over multilingual benefit would be obs erved if the BE group identified and discriminated the non native contras ts more accurately than the SE group, which in turn performed better than the AE monolingual group. Apart from these hypotheses, the results for the generalization test were expected to be fairly equal to or lesser than the posttest identification test res ults, since the successful generalization to a new voice would indicate the formation of new phonetic categories corresponding to the non native contrast. This prediction was based on the results from the previous training studies examining generalization of newly acquired contrasts to new voices (Logan et al 1991, 1993, Lively et al. 1994, Bradlow et al. 1997). The results of the current study do not show robust empirical evidence to support the hypotheses However, this may be due to the complex nature of the stimuli (4 pair of novel non native contrasts to acquire) as well as very limited time period for implicit perceptual training (6 sessions 35 minutes each). The subjects were required to perceive distinctions among eight target sounds, four of which were novel for the subjects [ ]. Acquiring all these

PAGE 103

103 four contrasts within the training period of only 3.5 hours (6 sessions of 35 minutes each) would have been a very daunting task. Perhaps, over a longer period of time, the robustness of the results will be seen. Therefore, to observe any robust effects in such limited training, the influence of the factors being explored would have had to be really strong. The hypothesis for multilingual benefit was support ed in one particular contrast ( l ateral non native contrast) out of the four contrasts used as stimuli. The evidence for multilingual benefit was consistent in the lateral contrast in the discrimination results as well as in all the subsequent reanalysis conducted. The results of the per ceptual assimilation test revealed effects of multilingual benefit prevalent in the bilingual groups. The statistical analysis of the proportion of change in assimilation patterns among the language groups was observed to be significant, with the BE group having a larger change in assimilation patterns after training as compared to the monolingual AE group but not significantly different from the bilingual SE group. The SE group revealed a marginal significant difference in the proportion change from that o f AE group. Therefore, although the evidence supporting the effect of multilingual benefit in facilitating the acquisition of non native contrasts in bilinguals is not robust enough, it is suggestive of presence of this factor. The hypothesis of feature ge neralization/feature productivity did not surface as a facilitating effect where multilingual benefit was not seen, such as a scenario where the BE group would perform better than the SE and AE groups with no group differences between SE and AE. The only i nstance which could be interpreted as perhaps an effect of feature generalization was the results of the nasal non native contrast in the case of the identification test. However, even in this case, the SE group was at borderline significance with the AE g roup, resulting in a very weak interpretation of the feature generalization hypothesis. In addition to

PAGE 104

104 that, an additive effect that was hypothesized to result from having previous linguistic experience with the retroflex feature in addition to multilingua l benefit was not observed in any area of testing. The perceptual assimilation test results along with requiring a descriptive approach allow us to make a priori prediction concerning the multilingual factor. It was expected that the changes in assimilatio n types before and after training, if any, w ould reveal effects of multilingual benefit ; that is, both bilingual groups would show a greater shift towards the assimilation types that clearly reflect learning of the non native contrasts than the monolingual group. The results of the study show support for this prediction. Group differences were observed between the bilingual groups when compared with the monolingual group, where the bilingual groups show a considerable decrease in SC (single category) and UU (un categoriable) assimilation types and an increase in the TC (two category) and CG (category goodness) assimilation types. This indicates that learning took place even during the very limited time period of training. The learn ing was seen more in biling uals than monolinguals. An overall shift in assimilation types from within category non discriminable types to more discriminable categorical assimilation types were observed across all language groups and all four contrasts. The PAM L2 model (Best & Tyler 2007) provides us with an approachable system of assimilation types that can be used to examine the perceptual assimilation performance before and after training. However, some caveats in the model restrict us from further exploring the direction of learna bility through these specific assimilation types. These are discussed in the next section in detail along with other unexpected results. Alternate Explanations If the results of the current study show support for certain hypotheses, there are other results and cross language group differences that are not so easily explained. This section

PAGE 105

105 provide s either possible explanations or interpretations of the various unexpected results found in the s tudy. Test of Generalization Results The results in the area of te st of generalization during the posttest phase of the training were contrary to the prediction. It was expected that the posttest scores of identification would be higher than or fairly equal to the generalization test scores as seen in previous literatur e. Moreover, if the explored factors were in fact effective in enhancing the acquisition of the non native contrasts, it was expected that the bilingual groups would show near equal test scores for posttest identification (trained on voice) and generalizat ion (novel voice) whereas the monolingual group would show a lower score in generalizing to a novel talker than scores attained with the trained on talker. However, the results of this study supported no such claims. On the contrary, there were no group di fferences between all three language groups. Moreover, all three language groups displayed generalization scores (novel talker) that were higher than the posttest scores (trained on talker). On analyzing the individual contrasts in order to tease out perha ps one particular contrast that may have been the easiest to discriminate, it was found that no one contrast was consistently being identified accurately by all speakers. There was a considerable variation in identification scores within groups and across contrasts. Therefore, one particular contrast could not be held responsible for these unexpected results. Another explanation for these surprising results could be the issue of baseline talker intelligibility issue. This inherent/inbuilt intelligibility of an individual talker is perhaps the factor affecting the results of the generalization test. Perhaps, the hyper articulated sounds produced by the novel talker made it easy for the all the subjects, across the language groups, to identify the stimuli more accurately than was the case with trained on talker. The term talker intelligibility can be explained in regards to the articulation of the sounds by individual talker. It

PAGE 106

106 a rticulated speech which enhances the intelligibility of the speech for the listener of the same language. On the listener, with same phonetic category inventory in the context of various kinds of noises. The acoustic/articulatory features are indicated clearly so as to be intelligible in adverse listening conditions. It may perhaps allow the talker to enhance the distinctions between contrasts in a phonological space (Uchanski 2000 ; Smiljanic and Bradlow 2005, refers to the normal rat e of speech that a talker will pro duce in ideal listening conditi ons, that is, not hyper Therefore, based on this interpretation, the results of the generalization test scores can be explained in the following way. Probably, the novel talker used in the t est of generalization provided the stimuli, of which certain sounds were clearly hyper articulated among the four speech contrasts, unlike the stimuli from the trained on talker used in the posttest. The hyper articulated sounds may have provided all the s ubjects with easy identification of those sounds resulting in better accuracy than with the trained on voice. With introduction to multiple talkers during training, the formation of new phonetic categories, which was the desired effect, to accommodate the distinctions between the new non native contrasts may have started among the subjects of the bilingual group. Therefore, with the baseline talker intelligibility varying considerably for the trained on talker and the novel talker, the subjects were able to perform slightly better with the new voice than the trained on voice. Although, no significant outcomes were observed for this test, the results are still suggestive of the fact that a direction towards development of robust phonetic categories correspond ing to the non native contrasts could be

PAGE 107

107 seen since all groups fared better in generalization scores. The monolingual groups scored equally well as they did with the trained on voice, like the bilingual groups. Contrary to hypothesis, it could be interpret ed as an effect of the enhanced intelligibility of the novel talker in increasing the performance level for the monolingual group. This interpretation of the results also reflects the robustness of the training program already observed in many high variabi lity training studies ( Logan et al. 1991 ; Lively et al. 1993 1994 ; Bradlow et al. 1997 ) Many perceptual training studies where multiple talkers are introduced during training he listeners as a confounding factor for the extreme va riations in the results (Lively et. al., 1993, 1994; Iverson et al ., 2005 ; Semiljanic and Bradlow 2007 etc.). However, this area of inherent talker intelligibility requires a deeper probe than a mere mention in the training studies as a confounding variable. This aspect of perceptual training method can be looked at exclusively in future research work which can contribute substantially in enhancing the efficacy and accuracy of the high variability perc eptual training method. Necessity for R eanalysis The design of the experiment was constructed with the assumption that the factors being explored w ould show robust effects within the individual bilingual groups. Therefore, a standard sample size of twenty subjects in each group seemed quite appropriate for the design. Additionally, having a number of contrasts in the stimuli was expected to provide the study with a range of perceptual data that could potentially contribute to the existing pool of explored c ontrasts for the cross language speech perception field and could be used for future work. On the contrary, the outcome of the existing design showed no effects of the feature generalization factor. Additionally, the effects of multilingual benefit showed only trends towards its presence and only in certain contrasts. No strongly suggestive support emerged

PAGE 108

108 through any of the testing measures for these factors that may influence the acquisition of the non native contrasts among bilinguals when compared with monolinguals. A further inspection seemed appropriate in order to examine the reasons for these varied unexpected results. Therefore, two reanalysis w ere conducted looking at two issues : inefficacy of the sample size and individual examination of the non native contrasts. Since the presence of the multilingual benefit factor was considered as merely suggestive within each bilingual group in the main results of the study, looking exclusively at the multilingual factor was considered as the next step. Combin ation of the two bilingual groups and comparison with the monolingual group revealed significant group differences implying that bilinguals had performed better at acquiring the non native contrasts than the monolingual groups. The pooling of bilinguals in to one group provided evidence for multilingual benefit as the bilingual group showed significantly better identification results than the monolingual group. The reanalysis by pooling the two groups together displayed a significant effect of multilingual b enefit that gave the individuals with two or more language systems an ability to extract relevant acoustic cues for the place distinctions of the non native contrasts and be able to learn to identify the sounds in these contrasts separately. However, this reanalysis revealed some caveats in the original design of the experiment, namely, the issue of sampling size. It appears that by increasing the sample size of the language groups, the effects of multilingual benefit become robust Additionally, a large de gree of individual variation within the language groups pointed towards the need for a larger sample size. The individual variability could have resulted from the uncontrolled factors such as the individual aptitude for learning. The power issue in this st udy br ought forth the fact that multilingual benefit effects are not dominant enough to be seen as a robust effect within a small sample of the population. In order to see significant group

PAGE 109

109 differences brought forth with multilingual benefit as the determi ning factor, a larger sampling size may be required, especially with very limited laboratory training period as in the current study. The other basis for an alternate reanalysis was examining all the non native contrasts individually. The reason behind thi s examination was to tease out any particular contrast whose results may be suppressing an effect of multilingual benefit The large variation in the perceptual performance on these contrasts pre and post training was another motivating factor to conduct a reanalysis. Upon a closer inspection of the four contrasts in question, the non native contrast that differed in both place and manner appeared to be most different from the rest of the stimuli. There were several reasons that emerged to exclude the rhot ic non native contrast from the dataset and reanalyze it. The reasons have been discussed in detail in chapter three and will be briefly restated here. Firstly, the results from the preliminary experiment showed that of the four non native contrasts, rhoti cs was the most easily discriminable by the monolingual American English group. Therefore, we see high perceptual performance scores for the rhotics contrast. Therefore, the results of this contrast may have suppressed the group difference significance not iceably. Secondly, the sounds [ ] of the rhotics contrast are distinguish ed in place as well as manner unlike the other three contrasts the sounds of which differ only in place distinction. This criterion perhaps provided the listeners with more of the s alient acoustic cues resulting in easy identification and consequently easier learning of this non native contrast which may have again been the reason for not showing significant group differences in the main result. Moreover, the phonetic definition of the Malayalam retroflex approximant has not reached consensus among the phoneticians who have examined the Malayalam consonant inventory. Lastly, the fact that

PAGE 110

110 Spanish English bilinguals will already have separate established phonetic categories for the al veolar tap, acquired through Spanish, and the alveolar approximant, acquired through American English, making this contrast the most easily discriminable for the Spanish English bilinguals. This would not have been the case with Bengali English bilinguals since they used the Indian English dialect which substitutes the alveolar tap with the American English approximant leaving the BE bilinguals as the only language group with no added advantage from this contrast. Together, these reasons motivated the exclu sion of the rhotic contrast from the analysis. Therefore, a reanalysis of the data set with only the place contrasts was conducted. The results of the only place contrast reanalysis, excluding the rhotic contrast results, for the identification test reveal ed significant group differences among the three language groups, as expected. On post hoc testing, the results of the lateral non native contrast were suggestive of the effects of multilingual benefit The bilingual groups performed equally well at identi fying the contrast. The BE group performed significantly better than the AE group with SE group showing marginal significance with the AE group. The results of AX discrimination test which had originally showed no significant group differences, revealed a marginal main effect for language groups. The results from this reanalysis were suggestive of the assumption that perhaps the results of the rhotic non native contrast were preventing the effects of multilingual benefit to surf ace as a robust effect. The r esults also shed light on the phenomenon that factors like multilingual benefit which enhance the perceptual performance or facilitate in the learnability of non native contrasts, may not be evident in a laboratory research with average difficulty contras ts, such as rhotics in this study In terms of inherent acoustic difficulty in perceiving the relevant cues of the contrast,

PAGE 111

111 studies have shown that spectral properties are less easily discriminable than the temporal properties when they are the distinguis hing acoustic cue within the contrast. Alternatively, in terms of previous linguistic experience/ native perceptual sound inventory, the PAM model of cross language speech perception (Best & Tyler, 2007) details that non native contrasts which are assimila ted to a single phonetic category as equally good exemplars or equally bad exemplars (SC type) or UU (no assimilation with less phonetic distance between them) type, will have a higher level of difficulty in discriminability. According to PAM, d uring the l earning process as well, the learner will exert more time in overcoming the difficulty level of this contrast and forming separate phonetic categories corresponding to the contrast. Therefore, only when the target new non native contrast is inherently at a higher level of difficulty will the effects of multilingual benefit emerge as robust. Thus, conducting reanalysis of the dataset gathered from this experiment revealed different perspectives of the factors explored in this study, which could be potential candidate hypothesis for future work. Caveats in Perceptual Assimil ation Analysis The results of the perceptual assimilation test were descriptive in nature. Reiterating the results, a n overall shift in assimilation patterns w as observed across the three l anguage groups and all four non native contrasts. A shift from assimilation types like SC (single category) or UU (uncategorizable) towards more categorical/ higher sensitivity for discrimination assimilation types like TC (two category) CG (category goo dness) or UC (uncategorizable categorizable) were observed. Noticeably, t hese results reflect the growth of the learning curve for all the three language groups since a decrease in SC assimilation types was consistently observed from pretest to posttest. S imilarly, an increase was observed in assimilation types like TC, CG and in some cas es UC from pretest to posttest.

PAGE 112

112 Although these learning shifts were seen for all of the language groups, a difference in the percentage of shift was noticeable between the monolingual group versus the bilingual groups. The bilingual groups, however, fared equally well, and their shifts in assimilation types were comparable. Therefore, in order to examine group differences statistically, an analysis of the proportion of chang e in the assimilation patterns was conducted which revealed significant main effect for language groups as well as significant effect for the multilingual factor (during reanalysis). The post hoc tests provided suggestive evidence for the multilingual bene fit factor. However, this analysis of proportion change concerns only the comparison of proportion of change in the assimilation type which in no way describes the direction of learning. In order to examine learnability among the groups in terms of percept ual assimilation a definitive ranking of the assimilation types needs to be established. For this study, five assimilation patterns (SC, UU, UC, CG and TC) were used to assess the perceptual performance of the learner groups. These assimilation types wer e adopted from the revised PAM model (Best & Tyler 2007) Since the current model of perceptual assimilation does not delve into the prediction of learning stages based on the assimilation patterns revealed in the perception of an experienced learner befo re and after training, the hypothesized learning continuum is proposed as an extension of the PAM model. T o assess the direction of learning in the perceptual assimilation mode a continuum of learning is postulated where SC (single category) and TC (Two ca tegory) assimilation type form the two extremes. According to PAM SC assimilation type at the initial exposure reflects poor discrimination of the given contrast since both sounds fall under one phonological category of the existing language of the learne r. The model predicts that the learner may not form a new phonological category at all where SC assimilation takes place In case of TC ( two category )

PAGE 113

113 assimilation at the initial stage of perception, the type reflect s of t he contrast. The learner already distinguishes the sounds of the contrast into two different phonological categories. Therefore, after becoming an experienced learner (in a classroom setting or natural environment) or at the post training stage (in a contr olled setting), as is the case in this study, if the learner maintains (pretest to posttest) the SC type of assimilation for a particular contrast, no learning has taken place. In case the learner shows TC assimilation before training and maintains the sam e level of assimilation post training again no learning has taken place since the learner is already maintaining a good distinction within the contrast. However if the learner progresses from SC type at the initial exposure stage to TC type at the post t raining stage, evidence of successful learning can be inferred This shift in assimilation pattern would provide evidence that the learner has been able to place the sounds into two different pho netic categories at the end of the training. In general, any shift in assimilation patterns towards TC during the course of training would evidence learning. Consequently t aking these assimilation types as the two extreme ends a learning continuum is proposed (Figure 4 1) Other assimilation types that are placed i n this continuum are CG (Category goodness), UC (Uncategorizable categorizable) and UU (both uncategorizable). The ranking of these assimilation types can be speculated as follows. According to PAM, CG is a within category assimilation, like SC type, and a llows poor to good discrimination. UU and UC are both between category assimilation patterns, similar to TC type, and may show moderate to excellent discrimination. Therefore, to determine direction of learning, a ranking similar to this can be assumed SC< CG
PAGE 114

114 However, this ra nking sequence is questionable The ranking of SC and TC assimilation patterns appear to be decisive on the continuum but this is not the case with CG, UU or UC Looking back at the PAM model for a more detailed description of these assimilation types, a d efinitive distinction between the CG UU and UC assimilation patterns seems missing According to PAM, CG assimilation type reflects the perception of a particular contrast as being within the same L1 phonological category but one phone being considered as the good exemplar of the category and the other as being the deviant exemplar of the same category During the learning process, the learner may form a new phon etic category for the perceived poor/deviant phone of the L1 phonological category but only gra dually. The good exemplar L2 phone may remain assimilated to the L1 phonological category This definition ranks the CG type above SC type and lower to the TC type in terms of learn ing Also, as a general assimilation pattern, CG is a within category t ype which restricts the placement of CG on the continuum closer to SC as related to TC. However, the assimilation type CG representing a within category assimilation pattern, which may equate to non learning type, may be seen with a different perspective. Based on the results of the current study, it is observed that if the learner provides variable goodness ratings (e.g. ratings 2 and 5 on the scale of 1 to 7) for the target contrast implying that he perceives the sounds as good and bad exemplar of the sam e phonetic category at the pre training level, and over the course of training shifts to perceiving the same target contrast consistently as two extremes of the ratings scale (at post training ratings 1 and 7 on the scale of 1 to 7), then perhaps successfu l learning does take place. However, according to PAM, the general assimilation pattern of CG type is still within category, that is, the learner still labels the modal response for each sound of the contrast as the same. Yet the learner consistently perce ives these sounds as

PAGE 115

115 different (one sound at the goodness rating of 1 (highly dissimilar to the modal response chosen from their native sound inventory) and the other sound at the goodness rating of 7(highly similar to the modal response)) at the post trai ning level. The problem here lies with the orthography of the language. Since this study was a controlled laboratory training experiment and learners were asked to use an open set from their language systems for perceptual assimilation task, the learners d id not have new orthographic representations for the newly formed phonetic categories to substitute and thus represent a TC type rather than a consistent CG type. However, in real life situation, this dilemma may not arise as a learner may use the orthogra phic representations corresponding to the phonetic categories of the target language while acquiring the non native contrast. In other scenarios, where orthography of the target language is not available or not being used, the learner will have lexical ite ms through which the phonemic distinction will be clear and thus, the learner can eventually form separate phonetic categories for these sounds and assimilate the contrast to a TC type. This may not be the case in controlled laboratory training settings wi th a limited time period which may result in subjects still using the same modal response but consistently discriminating the non native contrast. Therefore, in this light, CG types do show a trend towards learning. Another case of perceptual assimilation is that an assimilation pattern is termed UC i f only one of the L2 phones is perceptually assimilated with an L1 phonological category and considered a good exempla r of that category. According to PAM, t he discrimination of the speech contrasts would be e xcellent as the learner would always perceive the fully assimilated L2 phone separate from other less assimilated L2 phones. This places the UC type closer to the TC type on the continuum which means that having UC type at post training stage would show a trend towards learning. However, with PAM specified definitions, the ranking of UC type to CG

PAGE 116

116 type cannot be ascertained since CG type can allow for a wide range from poor to good discrimination of the L2 contrast Moreover, the UC assimilation type is n ot limited to just one scenario as explained in the PAM model Based on the results from this study it has been observed that t he learner provides various responses for categorization of the o ther L2 phone (the uncategorizable one), one of which may overlap with the categorization of the fully assimilated L2 phone T he discrimination then, may not be as clear with the minimally contrasting words in the target language. For example, the learner may assimilate one L2 phone consistently with L1 phonological cat assimilate the more deviant L2 phone to various other L1 In this scenario, UC assimilation type cannot be interpreted to successful learning since at least ten to fifteen percent of the times the L1 phonological category with which one L2 phone is fully assimilated, overlaps with the other uncategorizable L2 phone. Therefore, even though the learner consistently identifies one L2 phone with one L1 category, the identification of the mini mal contrast of the sounds is not learnt. Perhaps, in the real life situation, when the uncategorizable L2 phone is placed in minimal contrast with other sounds of L2, the learner may gradually form a new phonological category for it. Similarly, the case o f UU assimilation type is also quite ambiguous when it comes to the placement on the learning curve. The UU type represents assimilation where the learner does not assimilate the L2 phones to any one of his L1 phonological categories. PAM claims that the d iscrimination of this kind of speech contrast may depend on not just the comparative features of the given sound to its closest L1 counterpart but also its relationship to other phones in the same phonological space. If the L2 phones are assimilated to sim ilar set of L1 sounds then discrimination of this contrast is difficult for the learner since the acoustic proximity of the L2

PAGE 117

117 phones to each other and to the set of assimilated L1 phones will be less. On the other hand, if the L2 phones are assimilated to different set of L1 sounds the acoustic distance between the L2 sounds and the set of assimilated L1 sounds will be much more and therefore would result in easier discrimination by the learner. Thus in the case of UU type, the level of discrimination depe nds on the inherent acoustic distance between those two sounds as well as the L1 categories Again, the possibility of overlapping of L1 categories among the sets of sounds that are assimilated to the L2 contrast is immine nt in the UU type. Moreover, the listener may be providing more than one modal response for each category and still not be consistent with the goodness ratings. All these scenarios make it difficult to rank the UU assimilation type against the UC and CG ty pes. Following definitions specified in the PAM model, a ssimilation types like category goodness (CG) and the categorical (UC and UU) patterns cannot be allocated a ranking on the learning continuum decisively. If the general assimilation pattern is consid ered then CG (within category) has to be placed lower than UU and UC (between category). On the other hand, predictions of PAM allocate a very broad range of discrimination levels from poor to good for both CG and UU. Also, b oth UU and UC can possibly have lot of scenarios with varying assimilation results as discussed above Therefore, ranking these highly sensitive assimilation types on the learning continuum solely based on PAM definitions of the assimilation patterns is problematic. Based on the percept ual assimilation results in this study, a n attempt is made to speculate a learning continuum which shows the assimilation types representing various stages of learning L2/L3 by adults. Figure 4 1 shows the hypothetical learning continuum with five differen t assimilation types representing different types of learning.

PAGE 118

118 Figure 4 1. Hypothetical representation of the learning continuum displaying the ambiguity of ranking the UU, UC and CG assimilation types in order to assess the direction of learning. Th e SC and TC assimilation types form the two extreme ends of the continuum where SC type at the post training stage equates to no learning taken place and the TC type represents successful learning. The ranking of the CG, UU and UC assimilation types toward s learning cannot be decided without more empirical evidence supporting the ranking. Therefore, based on the results of this study, the ranking of these three assimilation types can only be speculated. Between CG and UU types, even though CG assimilation i s a within category pattern its goodness ratings of the deviant L2 phone exemplar reveal that the learner perceives the L2 phones apart. Also, the perceptual assimilation results show a drastic increase in CG assimilation types with consistently large good ness rating differences when pre training and post training level assimilations are compared. UU assimilation type is a between category assimilation pat tern which implies that the L2 phones are assimilated to different set of L1 phones. Other between cate gory assimilation patterns like TC denote successful learning but same cannot be applied for UU type. As discussed above, UU assimilation type can result from many scenarios which may not confirm as ev idence of learning. Therefore, CG can be ranked above U U assimilation type. UC assimilation type can also result from ma n y scenarios as discussed above. However, in this case, one L2 phone is fully assimilated to one of the L1 phonetic category which results in better discrimination of the L2 contrast. Moreov er, a shift towards UC assimilation type at the Direction of Learning SC UC? TC CG? UU?

PAGE 119

119 post training stage signifies that the learner has perhaps progressed from not distinguishing the L2 phones at all to being consistently assimilating one L2 phone to one L1 phonetic category. This indicates p rogression towards learning and thus places the UC type above UU type and more towards the right extreme of the learning continuum, that is, the TC extreme. The overall perceptual assimilation results also validate this ranking as they show a decrease in t he UU assimilation type and an increase in UC assimilation type at the post training stage. Between UC and CG assimilation types, no definitive ranking can be stated betwee n these two assimilation types since both the types show an increase at the post tra ining stage. Therefore, both these assimilation types can be positio ned anywhere between UU and TC. Thus, the ranking of assimilation types representing the learning continuum can be hypothesized as SC
PAGE 120

120 more sensitive to distinctions in auditory acoustic stimuli and being able to abstract the linguistically relevant information to contrast the speech sounds a nd develop a new phonetic category? Secondly, would effects of feature generalization be observed in performance of learners who may already have linguistic experience with a phonetic feature which is present in the new non native contrast as well? Lastly, what kinds of changes are seen in the assimilation patterns of non native contrasts p ost training? The results of this study contributed information as well as a new range of data set to the existing literature in cross language speech percepti on Previou s studies have examined many factors that contribute towards the perception of second language (L2) sounds and, in some cases, the subsequent establishment of L2 phonetic categories. Phonetic, phonemic and acoustic factors among others have been determined to influence the learning of L2 sounds (Tees and Werker 1984 ; Polka 1991 ; Best 1995 ; Flege 1995). However, learning a third language, unlike learning a second language, may be influenced by additional factors attributed to the presence of two language systems in a bilingual instead of one language system in a monolingual (Cenoz & Valencia, 1994; Ardeo, 2000; Munoz, 2000; Bild & Swain, 1989) This hypothesis was based on the studies in lexical processing and word learning by bilingual and monolingual ch ildren showing that bilingual children are more adept at doing metalinguistic tasks and develop ing cognitive benefits (Peal and Lambert 1962 ; Malakoff 1992 ; Klein 1995 ; Sanz 2000 etc.). N umerous other studies looking at cognitive processing at a lingui stic level have explored the concept of multilingual benefit in bilingual and multilingual children as well as adults which emphasize d the fact that bilinguals with near native proficiency in both languages show evidence of multilingual benefit (Bialystok 1992, 2001 2004 ; see also review in Jessner 2006). However, this concept of meta linguistic awareness is not yet fully explored in the field of cross language speech perception and the acquisition of new phonetic

PAGE 121

121 categories by bilingual learners of an L3. Th e results of the present study extend the research of metalinguistic ability to the field of cross language speech perception The study provides suggestive evidence that t he bilinguals function at a level different from monolinguals in reorganiz ing the information gathered from the speech stream in order to deal with the high demand task of acquiring new non native speech contrasts The results support the claims of previous studies on lexical processing with respect to the effects of multilingual b enefit seen in acquiring a third language by bilinguals. For instance, t he study by Klein (1995) showed that during lexicon acquisition of a language, multilinguals learned a higher number of lexical items than monolinguals. Klein based the explanation for this phenomenon on enhanced cognitive skills in multilinguals which helped them tease out the potential relevant data for resetting the particular parameter for the new language. Similarly in the present study, the results of the lateral non native contra st in particular showed that bilinguals, with or without previous experience with the retroflex feature, displayed better perceptual performance than the monolingual group. The study reflects the effect of multilingual benefit which may be further explored as future research in speech perception. The other factor that was explored in this study feature generalization, did not show any effect s The exploration of this factor was based on the assumption that features, an important set of information contain ed in the established phonetic category, play a crucial role in developing new phonetic categories (Nosofsky 1986, 1987; Kruschke 1992; Jusczyk 1989; Lively et al. 1993; Francis and Nusbaum 2002). Few studies have looked at the feature generalization phenomenon in the field of cross language speech perception. Polka (1992) and Harnsberger (1998) found results contrary to the expected results in this study that feature generalization will reveal better perceptual performance. Nevertheless, a study by Mc Allister, Flege & Piske (2002)

PAGE 122

122 found in the data from a few subjects that experience with the duration feature did facilitate so that native AE speakers performed better than native Spanish speakers who had no experience with duration feature The difference between this study and others was that the subjects were experienced learners of Swedish and used the target language often. Th eir study was suggestive of the hypo thesis of feature generalization. Therefore, one factor that could explain the previous studies conflicting results is that these studies (Polka, 1991; Harnsberger, 1998) examined only nave perceivers with speech perception experiment. The present study e mployed an extended approach which provid ed experience to the listeners through training the next step from the previous studies where feature generalization was tested on only nave listeners. The robustness of the high variability perceptual training pr ogram was assumed to be effective in providing sufficient linguistic exposure to the learners in order to bring forth any presence of the facilitating effects of these factors. The present study was unable to provide the missing link between the studies w here the feature generalization factor was explored but not found and the study which indirectly suggested the presence of feature generalization. Regardless of gaining from perceptual training with multiple talkers and high variability in stimuli tokens, Bengali English speakers were not able to reach a native both testing areas of identification and discrimination was not indicative of any robust effects of feature generalization. Al though, Bengali English speakers performed significantly better than the monolingual group of American English speakers, they were unable to display an additive effect over the metalinguistic ability found in the Spanish English speakers as well. Since bot h Bengali English speakers and Spanish English speakers performed near equally in most of the

PAGE 123

123 testing measures, the performance of a bilingual group (SE) without any previous experience with the retroflex feature negated any trace of the presence of featur e generalization that may have been possible. Thus, the present study confirmed the results of the earlier studies (Polka 1992 ; Harnsberger 1998 ; Tajima et al. 2008) that whether the listeners are nave or received limited implicit training, they are un able to ge neralize the phonetic features with which they have had previous linguistic experience Additionally, recent studies have shown that only phonemic experience/ language experience over time facilitates the perception of non native contrast and tha t simultaneous bilinguals acquire native like phonetic categories which are developed fully by adulthood (Silver 2007 ; Sundara et al., 2006, 2007 ). This may explain the results of the current study where no feature generalization was observed. The lack of extensive exposure of the stimuli to the language groups is one of the explanations provided for results showing no support for the feature generalization hypothesis. This possibility can be explored in future research work by elongating the span of train ing and providing extensive exposure of the non native contrast to subjects. Extensive training may result in observance of effects of language experience in this respect. The third focus of the study concern ed tual category inventory in the acquisition of non native contrasts from a third language. Several models have been developed to account for the influence of native perceptua l categories on the perception and acquisition of non native speech sounds, includi ng the Speech Learning Model (SLM) and, of greater interest for this project, the Perceptual Assimilation Model (PAM). PAM concerns the discriminability and learn ing of non native contrasts based on the relationship between the non native and native ca tegory inventories. These relationships have been encapsulated in several assimilation types which were adopted for this study. The results of the

PAGE 124

124 present study a ssess ed the extent to which initial p atterns of assimilation change following limited training experience and m easur ed the effects of the multilingual factor within the perceptual assimilation pattern shift. The results revealed a shift of perceptual assimilation types towards perceptual learning. On an average, the assimilation types SC and UU dec reased and the assimilation types TC, CG and UC increased from the pre training phase to post training phase. However, the relationship of learn ing could not be deciphered as the PAM model does not provide a definitive ranking of UU, UC (categorical as similations) and CG (within category assimilation) assimilation types with respect to learn ing In addition, the multilingual effects were also observed within the results of perceptual assimilation tests. The results were suggestive of the positive pe rceptual learning seen among bilinguals which appeared to be higher than the meager amount of perceptual learning that took place among the monolingual group. This study provides a substantial empirical dataset which can be used to explore the issue of lea rnability ranking among the assimilation types as future research work. Another area of future research that could be pursued would be to examine the rhotic segment (orthographic symbol: ) of Malayalam which lacks definitive phonetic description This s ound, as discussed in Chapter 2, has been presented in previous literature with various articulatory descriptions (Kumari 1972, Sreedha r 1972, Asher & Kumari 1997, Kalackel 1985, Raja 1960, Bright 1998, Krishnamurti, 2003) A clear phonetic description of this particular segment of Malayalam consonant inventory is required. One method that can be used to describe the features of this rhotic segment is acoustic analysis. It is observed that m ost of the descriptions affirm the feature of retroflexion in this sound Previous studies on Tamil liquids also report the occurrence of similar sound which includes features of retroflexion and central airflow (Narayanan et al 1996, 1999, McDonough & Joh nson 1997). This type of research of the

PAGE 125

125 Malayalam sound as well could provide a definitive phonetic description of the sound which could be compared to similar sound segment in other Dravidian languages and contribute to the knowledge of phonetic sounds o f the Malayalam language. Conclusion This dissertation extends the research on cross language speech perception and language acquisition to a new range of contrasts, different manners such as lateral, nasal, frica tive and rhotic with alveolar retroflex a s the place distinction. The study is successful in determining whether current theories and established claims of multilingual benefit in the field of lexical processing were generalizable to the area of cross language speech perception in regards to lear ning new non native contrasts similar to that of leaning of the lexicon. The results suggested that the effects of multilingual benefit were prevalent in the learning of novel contrasts with limited training period. However, the claim on the effects of fea ture generalization received no support in the results, therefore, confirm ing the findings of the earlier studies. In the area of cross language perceptual classification, the study provides a range of empirical datasets for assessing the cross language d ifferences in the perceptual assimilation patterns observed before and after training. It points to open questions of ranking of assimilation types in terms of learnability that may strengthen the predictability of perceptual assimilation in determining th e direction of learning. These concerns need to be addressed by the current models of perceptual assimilation. Beyond these general findings, the study also demonstrated the need for larger sampling size in order to get robust effects since the effects of multili ngual benefit observed were merely suggestive in nature and seen along with lot of individual variations. Overall, the results presented here lead to suggest a productive line of research for future work discussed in the above section. The caveats such as looking into the baseline talker intelligibility issue within training and testing; establishing learnability ranking of the assimilation patterns;

PAGE 126

126 examining with exclusive focus on the effects of multilingual benefit were revealed during the prog ress of this study. These caveats will provide substantial insight and enhance the productivity of the future research. In summary, the results supported only the hypotheses of multilingual benefit postulated in this study. The factor of multilingual bene fit is brought forth and explored within the realm of cross language speech perception field through the results of the study.

PAGE 127

127 APPENDIX A PRELIMINARY SCREENIN G Subject Information form Subject code (not to be filled by participant): Name: Age: Email: F irst language(s): Language(s) you speak with your parents/grandparents/siblings: Spoken proficiency in languages other than first language(s): Language classes that you might have taken in Middle/high school: gth of stay if more than 2 months): History of any hearing impairment: If participating in the study for credits specify name of course and Instructor:

PAGE 128

128 Screening for Spanish S poke n P roficiency This was a brief spoken test conducted at the time of scr eening the monolingual subjects as well as bilingual Spanish English subjects. The subjects were asked to say out loud, at normal speaking rate, two phrases which are considered as tongue twisters in Spanish. Then they were asked whether they were aware of the meaning of the phrase or certain words in them. This brief proficiency test was conducted to assess the level of spoken proficiency as well as lexical knowledge of the Spanish language. In case of monolinguals, if the candidates were not able to maint ain a normal speaking rate with less than 40% correct pronunciation (that is, they substituted English sounds like [ ] instead of [ ], [ ] instead of [ ], [ ] instead of [ ] and dropped etc. ), they could safely be consid ered under the monolingual language group, provided they had exposure to no other language. In case of self reported Bilingual candidates, high level of spoken proficiency was considered along with complete knowledge of lexical items and sentence structure Hay tres tristes tigres en un trigal. [ . ] IPA transcription El perro de san roque no tiene rabo. [ ] IPA transcription

PAGE 129

129 Fir st time interested participant email: Thanks for the interest in the study! You can see more details about the study at http://web.csd.ufl.edu/langbrain/Oth er_Research.htm#_Perception_training_study_2 Which of the language groups would you put yourself into? Before I go ahead and recruit you, you'll have to provide me with some information on your language background. At what age did you start learning you r second language, if any? What other languages have you been exposed to and for how long? (Please specify middle/high school, college, trips to places etc.) Also, did you work in any school or college projects involving a different language? Please let me know, if you've had any roommates or friends from whom you were exposed to a language other than English. What is your major, minor and year at UF? Also, where did you learn about the study? If you've any questions, please feel free to ask! Thank you! D ivya

PAGE 130

130 APPENDIX B SUBJECT INFORMATION Individual Subject Attributes Spanish English Speakers Name code Age Spanish Dialect First Language(s) Age of Exposure (L2) Linguistic background SE01 20 Venezuelan English, Spanish -Portuguese average proficienc y, French low proficiency SE02 23 Panamanian English, Spanish -exposure to Arabic SE03 20 Salvadorian Spanish 2yrs English, 2 years of French SE04 21 Cuban Spanish, English -N.A SE05 23 Argentinean Spanish 4yrs English, Portuguese average prof iciency SE06 19 Colombian Spanish 4yrs English, a little French SE07 20 Colombian English 6yrs Spanish SE08 20 Mexican Spanish, English -Introductory French and Portuguese SE09 19 Cuban Spanish 4yrs English SE10 20 Panamanian English, Spanish -N.A SE11 22 Cuban English, Spanish -N.A SE12 19 Salvadorian Spanish, English -N.A SE13 20 Nicaraguan Spanish, English -N.A SE14 19 Colombian Spanish, English -2 years of French (Middle school) SE15 19 Mexican English, Spanish -a little F rench SE16 20 Mexican Spanish 4yrs English SE18 19 Cuban Spanish 3 yrs English SE19 21 Mexican Spanish, English -N.A SE20 21 Colombian English 2yrs Spanish SE21 20 Nicaraguan Spanish, English -N.A

PAGE 131

131 Bengali English Speakers Name Code Age Cit y, State First Language(s) Age of exposure (L2) Linguistic Background BE01 28 Kolkota, West Bengal Bengali 8 yrs English, Hindi BE02 24 Kolkota, West Bengal Bengali 6yrs English, a little Hindi BE03 27 Kolkota, West Bengal Bengali 3yrs English, formal e ducation in Hindi BE04 25 Kolkota, West Bengal Bengali, English -English, Hindi, Sanskrit BE05 30 Kolkota, West Bengal Bengali 3yrs English, formal education in Hindi BE06 26 Kolkota, West Bengal Bengali 6yrs English, a little Hindi BE07 32 Kolkota, West Bengal Bengali 8yrs English, Hindi BE08 28 Kolkota, West Bengal Bengali, English -English, Hindi, exposure to Assamese and Gujarati BE09 32 Kolkota, West Bengal Bengali 8yrs English, a little Hindi BE10 24 Kolkota, West Bengal Bengali 2 yrs Engli sh, Hindi, exposure to Urdu and Kannada (3 mnths) BE11 24 Kolkota, West Bengal Bengali 8yrs English, formal education in Hindi BE12 24 Kolkota, West Bengal Bengali, Hindi -English, formal education in Hindi, exposed to Telugu (6 months) BE13 25 Kolkot a, West Bengal Bengali 6yrs English, a little Hindi BE14 27 Chandan Nagar, West Bengal Bengali 8yrs English, Hindi, exposure to Kannada (20 months) BE15 27 Chakda, West Bengal Bengali 3yrs English, a little Hindi and Nepali BE16 29 Barddhaman West Ben gal Bengali 5yrs English, Hindi, Sanskrit, German, a little Urdu BE17 28 Dhaka, Bangladesh Bengali 12yrs English, a little Hindi, Assamese and Oriya BE18 30 Kolkota, West Bengal Bengali 3yrs English, a little Hindi and Assamese BE19 28 Kolkota, West Be ngal Bengali 6yrs English, Assamese, a little Hindi BE20 32 Kolkota, West Bengal Bengali 8yrs English, Hindi

PAGE 132

132 American English Speakers Name Code Age First Language(s) Linguistic Background AE01 23 English Spanish Middle/High school AE02 20 English 2 years of Spanish AE03 19 English Spanish Middle/High school AE04 23 English a little Spanish AE05 19 English Introductory French and Spanish AE07 20 English a little Spanish and French AE08 19 English a little Spanish AE09 22 English 2 years of L atin, exposed to Spanish AE10 19 English 1 year of Latin AE11 19 English 2 years of Spanish(High school), 2years of Latin (Middle School) AE12 20 English 1 year of Spanish (Middle School) AE13 20 English 1 year of Spanish (Middle School) AE14 19 Engli sh 3 years of Spanish (Middle/High school) AE15 19 English 3 years of Spanish (High school) AE16 22 English a little Spanish and Italian AE17 19 English a little French, exposure to Spanish AE18 20 English 2 years of Spanish (Middle School) AE20 21 En glish 2 years of Spanish (High School), 1 year of L atin AE21 22 English 2 years of Spanish (Middle School) AE22 19 English 1 year of Spanish (Middle School)

PAGE 133

133 APPENDIX C EXPERIMENT INSTRUCTI ON SHEET Instructions given to subjects on the first day of th e experiment (Pretest Phase): Session 1 Instructions Welcome to the Linguistics Laboratory. We appreciate your participation in this experiment and also hope you will find it interesting. You will be participating in a study of the perception training of new sounds (consonants will carry on for 8 sessions Today is your first session. In this session, you will have to do four tasks. Task 1 Familiarization: In this task, you will listen to consonants from another language. Each consonant sound sound, regardless of the vowels, is associated with a symbol. You will hear a wo do is learn to associate the consonant sound with the symbol. This may be difficult initially as the sounds and symbols are from another l anguage. This will take no more than 6 minutes. Task 2 Identification task: ou Again, use the knowledge gained from the familiarization task. Since some of the sounds will be new to you, they may be difficult for you to tell apart

PAGE 134

134 which consonant symbol it is, just make the best choice possible and move on to the next sound. This will take about 15 20 minutes. Task 3 Discrimination task: In this test, you will hear two words in a row, se parated by a short pause. These words are from another language, and each word consists of a consonant preceded and followed by a vowel sound. Your task is to determine whether the two consonants are the same consonant or different consonants. If the two w For example, you might hear something like the following: ala (pause) ana In this example, relatively easy example. However, in this test, you will be hearing words from anot her language. Some of these words may be difficult for you to tell apart. However, even if you are not entirely sure if the consonants are the same or different, just make the best choice possible and move on to the next set of words. Task 4 Instructions for Perceptual Assimilation task: In this task, you will listen to words from another language similar to previous tasks. You must choose the best possible sound from your native language that closely corresponds to the sound given. Write the letter (or letters) best representing the speech sound on the sheet provided. Then compare the non native sound with closely related sound from your own language. On the scale from 1 7 circle the number that you feel is appropriate in comparing the similarity of th e non native sound with that of your own language sound.

PAGE 135

135 on your answer sheet and then scale the similarity of the two sounds. How similar are they? If 7 other options on the scale as well depending on your judgment. 1 2 3 4 5 6 7 Very different exactly same This is a paper responses on the sheet provided.

PAGE 136

136 AP PENDIX D PERCEPTUAL ASSIMILAT ION RESPONSE SHEET Note: This is a sample page of the perceptual assimilation response sheet. In all, 80 responses were elicited in this task. Name: Subject Code: Task: Serial No. Letter symbol Goodness Rating (1 7) 1 1 2 3 4 5 6 7 Very different exactly same 2 1 2 3 4 5 6 7 Very different exactly same 3 1 2 3 4 5 6 7 Very different exactly same 4 1 2 3 4 5 6 7 Very different exactly same 5 1 2 3 4 5 6 7 Very different exactly same 6 1 2 3 4 5 6 7 Very different exactly same 7 1 2 3 4 5 6 7 Very different exactly same 8 1 2 3 4 5 6 7 Very different exactly same 9 1 2 3 4 5 6 7 Very different exactly same 10 1 2 3 4 5 6 7 Very different exactly same 11 1 2 3 4 5 6 7 Very different exactly same 12 1 2 3 4 5 6 7 Very different exactly same 13 1 2 3 4 5 6 7 Very different exac tly same

PAGE 137

137 APPENDIX E RESULTS IN TABULAR F ORMAT Table E 1. Mean percent identification scores averaged over all four contrasts at the pretest and the posttest level. Language groups Pretest scores Posttest scores AE 51 63 BE 61 80 SE 56 74 Table E 2 averaged over all four contrasts at the pretest and posttest level. Language groups Pretest scores Posttest scores AE 0.85 1.3 BE 1.17 1.94 SE 0.95 1.64 Table E 3. M ean percent identification scores of cont rasts spoken by trained on t alker (posttest) and new talker (generalization test). Language groups Posttest scores Generalization scores AE 63 64 BE 80 82 SE 74 75 Table E 4. Mean percent scores of consonant identification test for all four contr asts by the bilingual groups (BE and SE) and the monolingual group (AE) at the pretest and the posttest levels. Language groups Laterals Nasals Fricatives Rhotics Pretest Posttest Pretest Posttest Pretest Posttest Pretest Posttest AE 47 55 51 59 47 57 59 82 BE 65 90 59 80 52 62 67 87 SE 56 74 54 68 46 67 67 88 Table E 5. groups (BE and SE) and the monolingual group (AE) at the pretest and the posttest lev els. Language groups Laterals Nasals Fricatives Rhotics Pretest Posttest Pretest Posttest Pretest Posttest Pretest Posttest AE 1.00 1.04 0.47 0.76 0.63 1.10 1.31 2.29 BE 1.58 2.77 0.90 1.38 0.99 1.45 1.21 2.17 SE 0.86 1.75 0.49 0.73 0.66 1.05 1.81 3. 04

PAGE 138

138 Table E 6. Reanalysis results: Mean percent identification as well as AX discrimination scores averaged over all four contrasts at the pretest and the posttest level for the bilingual groups (BE and SE) combined versus the monolingual group (AE). Lang uage factor ID Test Score AX Discrimination score Pretest Posttest Pretest Posttest Monolingual 51 63 0.85 1.06 Bilingual 58 77 1.3 1.79 Table E 7. Mean percent correct response (averaged over contrasts) for identification of the three place contras t s (laterals, nasals, fricatives) by bilingual (BE and SE) and monolingual (AE) groups at the pretest and posttest levels. Language groups ID Test Score AX Discrimination score Pretest Posttest Pretest Posttest AE 48 57 0.7 0.97 BE 59 77 1.15 1.87 SE 52 69 0.67 1.18

PAGE 139

139 LIST OF REFERENCES Agresti, Alan & Finlay, Barbara. (1997). Statistical methods for social sciences (3 rd Ed.). Prentice Hall. Aakhane Yamada, R., Tohkura, Y., Bradlow, A.r., & Pisoni, D.B. (1996). Does training in speech perception mod ify speech production? Spoken Language, 1996, ICSLP 96, Proceedings., Fourth International conference on. Vol. 2, pp. 606 609. Asher, R.E. & Kumari, T.C. (1997) Malayalam (Descriptive Grammars) London & New York: Routledge, 1997. Pp. xxvi, 491. Aiyar, L. V. R. (1972). Linguistics, 1(2) 10 18. Best, C T., McRoberts, G. W. & Sithole N.M. (1988) Examination of perceptual reorganization for nonnative speech contrasts: Zulu click discrimination by E nglish speaking adults and infants Journal of Experimental psychology: Human perception and Performance 14 ( 3 ) 345 60. Best, C. T., and Strange, W. ( 199 2). Effects of phonological and phonetic factors on cross language perception on approximants. J. P honetics 2 0 305 330. Best, C. T. ( 1994) The emergence of native language phonological influences in infants: a perceptual assimilation model, in The Development of Speech Perception: The Transition from Speech Sounds to Spoken Word s, edited by H. C. Nusb aum ,MIT, Cambridge, MA. Best, C. T. ( 199 5). A direct realist perspective on cross language speech perception, in Speech Perception and Linguistic Experience: Theoretical and Methodological Issues in Cross language Speech Researc h, edited by W. Strange ,Yo rk, Timonium, MD, 167 200. Best, C. T and Tyler, M. D. (2007). Nonnative and second language speech perception: Commonalities and complementarities i n Language Experience in Second Language Speech Learning: In Honor of James E. Fleg e edited by Ocke Schwen Bohn and Murray J. Munro, John Benjamins publishing. Bhattacharya, Krishna. (1988). Bengali Phonetic Reader Central Insitute of Indian Languages, Mysore. Bialystok, E. (1992) Cognitive Processing in Bi linguals Edited by R. Harris, Amsterdam 501 514. Bialystok, E. (2001). Annual review of Applied Linguistics 21 169 81. Bialystok E., Craik, F., Klein, R. & Viswanathan, M. (2004) and cognitive Psychology and Aging 19 (2) 290 303.

PAGE 140

140 Bild, E.R. and Swain, M. (1989). Minority language students in a French immersion Development, 10 255 274. Bradlow, A.R., Pisoni, D.B., Aakhane Yamada, R, & Tohkura,Y. (1997). Training Japanese listeners to i dentify English /r/ and /l/: IV s ome effects of perceptual learning on speech production Soc Am. 101 (4). Bright, W (1998). The D ravidian S cripts. In T he D ravidian L anguages edited by Sanford S teever. Routledge publications. 40 74. Cenoz, J. and Valencia, J. F. (1994). 15 195 207 Cenoz, J. (2001). The effect of linguistic distance, L2 status and age on cross linguistic influence in third language acquisition In Cross Linguistic Influence in Third LanguageAcquisition: Psycholinguistc Perspectives, edited b y Cenoz, Hufeison & Jes sner. 8 20. Cenoz, J. (2003). The International Journal of Bilingualism, 7 (1) pp. 71 87. Costa, A. (2004 ) Speech production in Bilinguals In The Handbook of Bilingualism, edited b y T. K. Bhatia and W. C. Ritchie Oxford: Blackwell publishing. 201 223 Chatterji, S K. (1921) Bengali Phonetics Bulletin of the School of Oriental Studies, University of London, Vol. 2 ( 1 ) 1 25. Cambridge University Press. De Bot, K. (1992). A Bilingual Production Model: Levelt's "Speaking" Model Approach Applied Linguistics, 13 (1):1 24. Dalby, J. and Kewley Port, Diane. (1999). 16 ( 3) 425 444. E nomoto, K. (1994). Linguistics, 5 15 29. Ferguson, C.A. and Chowdhary, M (1960). The Phonemes of Bengali Language 36 ( 1 ) 22 59. Flege, J. (1998 ) The phonetic study of bilingualism 35 19 28 Flege, J., MacKay, I., and Piske, T. (2002) Assessing bilingual dominance Applied Psycholinguistics, 23 567 598. Flege, J.E., Schirru, C. and MacKay, I.R.A. (2003). Interaction between the native and second language phoneti c subsystems Speech Communication 40 467 91

PAGE 141

141 Flege, J.E. (1995). Second language speech learning: Theory, findings and problems Speech Perception and Linguistic Experience: Issues in Cross Language research edited by W. Strange 233 277. Francis, A. L. and Nusbaum, H. C. (2002). Selective attention and the acquisition of new phonetic categories Journal of experiment al psychology: human perception and performance. 28 ( 2 ) 349 366. Gibson, E. J. (1991) An Odyssey in Learning and Perception. Cambridg e, MA: Bradford Books MIT Press Gibson, J. J. (1966) The Senses Considered as Perceptual Systems. Boston, MA: Houghton Mifflin. Gibson, J. J. (1979 ) The Ecological Approach to Visual Perception. Boston, MA: Houghton Mifflin. Gonzalez Ardeo, J. (2000) Engineering students and ESP in the Basque Country: SLA versus Looking beyond second language acquisition: Studies in tri and multilingualism pp 75 95, Stauffenburg. Grosjean, F. (1998). Studying Bil inguals: methodological and conceptual issues Bilingualism: language and cognition, 1 (2), 131 149. Grosjean, F., Li, P., Munte, T.F. and Rodriguez Fornells, A. (2003). Imaging Bilinguals: when the neurosciences meet the language sciences Bilingualism : Language and Cognition 6 (2) 159 165. Grosjean, F (1992) nother view of bilingualism In Cognitive processing in bilinguals edited by R J Harris E lsevier S cience publishers 51 62. Haldar, G. (1986). A Comparative Grammar of East Bengali Dialects Calcutta: Puthipatra. Halle, P. A., B est, C. T. and Levitt, A. (1999). Phonetic vs. phonological in fl uences on French listeners: perception of American English approximants Journal of Phonetics 27 281 306. Hammarberg, B. (2001). Roles of L1 and L 2 in L3 Production and Acquisition In Cross Linguistic Influence in Third Language Acquisition: Psycholinguistc Perspectives, edited by Cenoz, Hufeison & Jessner, 21 41. Herdina, P. and Jessner, U. (2002). A dynamic model of Multilingualism Clevedon: Mu ltilingual Matters. Hoffmann, C. (2001). Towards a description of trilingual competence International journal of bilingualism. 5 ( 1 ) 1 17.

PAGE 142

142 Hammond, R.M. (2001). The Sounds of Spanish: Analysis and Application Cascadilla Press. 2001. Harnsberger J.D. ( 1998) native nasal contrasts: A cross linguistic Iverson, P., Hazan, V. and Bannister, K. (2005) manipulations: A comparison of methods for teaching English /r/ /l/ to Japanese adults J. Acoust. Soc. Am. 118 (5) 3267 78. Jamieson, D., and Morosan, D. (1986) "Training non native speech contrasts in adults: Acquisition of the English / d / / / contrast by francophones," Pe rcept. Psychophys 4 0 205 215. Jamieson, D.G. & Morosan, D.E. (1989) Training new, non native speech contrasts: a comparison of the prototype and perceptual fading techniques. Canadian Journal of Psychology, 43 (1), 88 96. Jessner U. (2006) Linguistic Awareness in Multili nguals: English as a Third Language Edinburgh University Press. Jusczyk P (1989). "Developing phonologic a l University. Kalackel, T. (1985). A contrastive Analysis of the Phonological Systems of English and Malayalam, Ph.D. Thesis, Northern Illinois University. Klein, Elaine C. (1995). Second versus third language acquisition: is there a difference? Language Learning. 45 ( 3 ) 419 468. Kostic, D. and Das, R. (1972). A Short Outline of Bengali Phonetics, Calcutta: Statistical Publishing Society. Kroll, J. & Sunderman, G. (2003). Cognitive processes in second language learners and bilinguals: the development of lexical and conceptual representations In The Handbook of Second language Acquisition, edited by Doughty & Long, Oxford: Blackwell Publishing 104 129. Krish namurti, Bhadriraju. (2003). The Dravidian Language, Cambridge U niversity P ress 2003. Kirsner K ., S mith, M C ., L ockhart R S., K ing, M L. and J ain M (1984). The bilingual lexicon: language specific units in an integrated network Journal of verbal learning and verbal behavior, 23 519 539. Krashen, S (1981) S econd L anguage A cquisition and S econd L anguage L earning Oxford: P ergamon press.

PAGE 143

143 Kroll, J F., and T okowicz, N (2005). Models of bilingual representation and processing: looking back and to the future In H andbook of bilingualism: psycholinguistic approache s, edited by J F. Kroll and A. M. B. Degroot, New Y o rk: O xford Un iversity P ress 531 553. Kruschke, J. (1992). ALCOVE: An exemplar based connectionist model of category learning Psychol. Rev. 90 22 44. Kuhl, P. K. (1991) prototypes 50 93 107. Kuhl, P. K., Williams, K. A., Lacerda, F., Stevens, K. N., and Lindblom, B. (1992) "Linguistic experience alters phonetic perception in infants by 6 months of age," Science 255 606 6 08. Kuhl, P.K. and Iverson, P. (1995). Speech perception and linguistic experience: Issues in cross language research edited by W. Strange, 121 154. Kumari, B. S. (1972). Malayalam Phonetic Re ader. Mysore: Central Institute of Indian Languages. Lively, S.E., Logan, J.S. and Pisoni, D.B. ( 1993 ) Training Japanese listeners to identify English /r/ and /l/. II: The role of phonetic environment and talker variability in learning new perceptual cat egories Journal of acoustical society of America, 94 (3) 1242 55. Lively, S.E and Pisoni, D.B. ( 1994 ) Training Japanese listeners to identify English /r/ and /l/. III: Long term retention of new phonetic categories Journal of acoustical society of Am erica, 96 (4) 2076 20 87. Logan, J.S., Lively, S.E. & Pisoni, D.B. 1991. Training Japanese listeners to identify English /r/ and /l/: A first report. Journal of Acoustical society of America, 89 (2), pp. 874 886. Ladefoged, P. and Maddieson, I. ( 1996 ) The Blackwell Publishers. Ladefoged, P. and Bhaskararao, Peri. ( 1983 ) Non quantal aspects of consonant production: A study of retroflex consonants University of California Working Papers in Phonetics 57 11 25. Ladefoged, P. ( 2 005 ) Vowels and Consonants: an introduction to the sounds of languages Blackwell Publishing. Ladefoged, P. ( 1999 ) English In Handbook of the International Phonetic Association Cambridge Univeristy Press. Lindau, Mona. ( 1985 ) The Story of /r/ i n Phonetic Linguistics: Essays in honor of Peter Ladefoged edited by V A. Fromkin Academic Press, Inc.

PAGE 144

144 Major, R. and Kim, E. (1996). Language Learning 46 ( 3 ) 465 496. Malakoff, M. (1992) Translation Abilit y: A Natural Bilingual and Metalinguistic Skill i n Cognitive Processing in Bilinguals edited by R. Harris Amsterdam: North Holland, 515 30. McAllister, R., Flege, J.E. and Piske,T. (2002) native speaker s of Spanish, English and Estonian Journal of Phonetics 30 229 258. McDonough, J. and Johnson, K. (1997) Asso ciation, 27, 1 2. Mohanan, K. P. and Mohanan, T. (1984). 15 (4) 575 602. Mun In J. Cenoz and U. J essner (eds.) English in Europe: The acquisition of a third language pp. 157 178, Clevedon: Multilingual Matters. Narayanan, S., Kaun, A., Byrd, D., Ladefoged, P. and Alwan, A. (1996) ICSLP, Philadelphia, PA. Narayanan, S., Byrd, D. and Kaun, A. (1999) Neri, A., Cucchiarini, C., Strik, H. and Boves, L. ( 2002 ) The pedagogy technology interface in computer assisted pron unciation training Computer Assisted Language Learning, 15 (5) 441 467. Nosofsky, R. M. (1986). "Attention, similarity, and the identification categorization relationship," J. Exp. Psychol. Gen., 115 39 57. Nosofsky,R M. (1987). "Attention and learning processes i n the identif i cation and categorization o f integral stimuli," J. Exp. Psychol. Learn. Mem. Cogn 15 700 708. Olive, J.P., Greenwood, A. and Coleman, J. ( 1993 ) Acoustics of American English Speech: a dynamic approach New York: Springer Verlag. Peal, E. and Lambert, W. (1962) Psychological Monographs 76 1 23. Polka, L. ( 1991 ) Cross Language speech perception in adults: Phonemic, Phonetic and acoustic contributions Journal of Acoustical Socie ty of America 89 (6 ) Polka, L. (1992) Characterizing the influence of native language experience on adult speech perception Perception & Psychophysics 52 (1) 37 52.

PAGE 145

145 Pruitt, J.S, Jenkins, J.J. and Strange, W. ( 2006 ) Training the perception of Hindi dental and retroflex stops by native speakers of English and Japanese Journal of Acoustical s ociety of America. 119 (3) 21 90 96. Ray, P. S., Hai, M. A., and Ray, L. (1966). Beng ali Language Handbook, Washington D. C., Center for Applied Linguistics. Roach, P. ( 1983 ) English phonetics and phonology Cambridge: Cambridge University Press Sanz, C. ( 2000 ) Bilingual education enhances third language acquisition: evidence from Cata lonia Applied Psycholinguistics. 21 23 44 Silver, J. (2007) Smiljanic, R. and Bradlow, A. (2007) ICPhS XVI, Saarbrucken 6 10 August 2007, 661 664. Smiljanic, R. and Bradlow, A. (2005). Production and perception of clear speech in Croatian and English J.Acoust Soc of Am 118 (3) 1677 1688. Sreedhar, M. V. (1972). of Dravidian Linguistics 1.2 100 125. Strange, W. and Dittmann, S. ( 1984 ) Effects of discrimination training on the perception of /r l/ by Japanese adults learning English Percept Psychophys. 36 (2) 131 1 45. Strange, W., Polka, L. and Aguilar, M.C. ( 1989 ) Effects of auditory and phonetic training on dental contrasts Journal of Acoustical Society o f America, 86 ( S1 ) S101. Strange, W. (1995). Speech Perception and Linguistic Experience: issues in cross language research Baltimore: York Press. Subramoniam, V. I. (1973). Dravidian Lingui stic Conference III, Dharwad, India. Sundara, M., Polka, L. and Genesee, F. (2006) experience facilitate discrimination of /d 100 369 388. Sundara, M. and Polka, L. (2007) D iscrimination of coronal stops by bilingual adults: The 106 234 258 Tajima, K., Kato, H., Rothwell, A., Akahane Yamada, R., Munhall, K. G. (2008) English listeners to perceive phonemic len 123 (1) 397 413.

PAGE 146

146 Tees, R. C. and Werker, J. F. (1984) ability to discriminate non 38 579 590. Tung a, S. S. (1995). Bengali and Other Related dialects of South Assam New Delhi: Mittal Publications. Uchanski, R. M. (2000). In T he Handbook of Speech Perception, e dited by D.B. Pisoni and R.E. Remez, Blackwell Publishing 207 235. Werker, J. F., Gilbert, J. H. V ., Humphrey, K., and Tees, R. C. (1981) of cross 52 349 353. Werker, J.F. and Tees, R.C. ( 1984 ) Phonemic and phonetic factors in adult cross language sp eech perception Journal of Acoustical Society of America 75 ( 1 )

PAGE 147

147 BIOGRAPHICAL SKETCH Divya Verma Gogoi received her Ph.D. in l inguistics from the University of Florida in 2010 During her time at UF, she worked as a teaching assistant for the Departmen t of Linguistics and worked as a research assistant to Dr. Caroline Wilts hire. Her doctoral degree was supported by a grant from Language Learning: A Journal of Research in Language Studies. Before enrolling as a grad uate student at UF, she received her B. A and then M.A in English from the Panjab University Chandigarh, India in 1999 and 2001 respectively. Thereafter, s he completed her M.Phil in Linguistics at the Central Institute of English and Foreign Languages, India (now The English and Foreign Languag es University) in 2005 with specialization in Phonetics and Phonology. Her research interests are Phonetics and second/third language acquisition; m ore specifically acoustic phonetics and cross language speech perception and learning