Estimated potato stocks and production

MISSING IMAGE

Material Information

Title:
Estimated potato stocks and production
Running title:
Potato stocks
Physical Description:
1 issue : ; 28 cm.
Language:
English
Creator:
United States -- Crop Reporting Board
Publisher:
Crop Reporting Board
Place of Publication:
Washington, D.C
Publication Date:
Frequency:
monthly
regular

Subjects

Subjects / Keywords:
Potatoes -- Statistics -- Periodicals -- United States   ( lcsh )
Genre:
serial   ( sobekcm )
federal government publication   ( marcgt )
statistics   ( marcgt )
periodical   ( marcgt )

Notes

Statement of Responsibility:
Crop Reporting Board, Statistical Reporting Service, U.S. Department of Agriculture.
Dates or Sequential Designation:
Dec. 15, 1981.
General Note:
Title from caption.
General Note:
Sole issue.
General Note:
"Pot 1-2"-- caption

Record Information

Source Institution:
University of Florida
Rights Management:
All applicable rights reserved by the source institution and holding location.
Resource Identifier:
aleph - 001278190
notis - AGC8847
oclc - 08561373
lccn - 2010229699
Classification:
lcc - HD9235.P8 P67
System ID:
AA00008928:00001

Related Items

Preceded by:
Potato stocks (United States. Crop Reporting Board)
Succeeded by:
Estimated potato stocks


This item is only available as the following downloads:


Full Text

ESTIMATED


AND
Released:


POTATO


PRODUCTION


December


3:00


1981


P.M.


FEB1 i


.F.A.


- Univ.


orida


~.(


I


Statistical Reporting
Service
U.S. Department
of Agriculture
Washington, D.C.
20250


POTATO


STOCK


PERCENT


FROM


YEAR


EARLIER


Pot


percent
total s
percent


*


ato st
million
above


tock


;ock


metric
December


hand


White


of D
tons)


)ecember
in the


1980


percent


1981


major


percent


major


are
fall


State


Reds.


estimated
producing
s than two


percent


million


State
years


were


ussets,


three


percent
Maine t


above


otal


Eastern


a year al
.3 million


States,


cwt,


estimated


percent t
3 percent


holdings


elow D
from a


ecember 1
year ago.


million


1979


cwt,


Stock


(Comment


continued on


page


FALL


POTATO


PRODUCTION


STOCKS


NIL CUT


I i~Y~ W'IA 1W UOfl 1070 IWJO I 'u-I


STOCKS


Crop
Reporting
Board


I




NOTE:


-Stocks are defined as the quantity reman- --ng in storage for-all purposes---
Stocks are defined as the quantity remaining in storage for all purposes


and uses,


including shrinkage and waste and other


the date of each report.


Sales


of fall


losses that occur after


potatoes for all purposes generally


account for about 90 percent of the total


loss and home use


account for the remaining


fall production.
g 10 percent.


Shrinkage and


TABLE


AND MAY 1


FALL POTATOES:


DECEMBER 1


, JANUARY 1, FEBRUARY 1, MARCH 1, APRIL


TOTAL STOCKS, AND COMPARABLE PRODUCTION FOR 15 STATES ESTIMATING POTATO STOCKS 1/


--------- flee----- ------------- ------------- --- --- ------em---- ---- a ----- ------s--a--- -----s-s--s--a--- -


: PRODUCTION


DEC 1


JAN 1


FEB 1


MAR 1


APR 1


MAY 1


- a a ------ ----- as sea mesa-- ---- ------ a e ----------------- ------- ------- -a--ease-a ------- -


1,000 CWT


258,613
242,390
247,686
278,679
266,422
294,978
295,421
311,981
285,060
255,691
281,375


172,550
155,190
154,890
183,850
181,720
198,630
202,550
219,850
200,820
171,730
187,790


148,600
132,050
131,600
160,100
156,220
172,230
175,300
193,520
176,020
146,610


122,350
105,690
105,360
131,480
129,710
142,030
147,930
162,980
147,910
121.565


97,150
82,510
80,460
102,963
102,850
112,830
119,850
132,570
121,720
97,280


57,860
55,465
75,217
70,970
81,130
88,680
99,250
92,550
72,460


64,830
59,535
44,446


SEE TABLE


FOR LIST OF 15 STATES.


BEGINNING WITH


SUMMER PRODUCTION FOR N Y


-LI


WIS AND WASH HAS


BEEN CLASSIFIED AS FALL.


TABLE


. POTATOES


USED FOR PROCESSING 1


, SEVEN STATES,


1980 AND 1981


CROPS


STATE


: STORAGE
: SEASON


: TO
: DEC 1


: JAN 1


: FEB 1


: MAR 1l


: APR I


* MAYl


ENTIRE
SEASON


----- ------- ----- --------------- ------- ------ ------ C C --------- --- ------------- ----- -------- -
S
U


IDAHO AND MALHEUR
CO.. OREG


1980-81
1981-82


14,460
16,720


18.710


23,230


28.110


33.570


39,210


52,720


MAINE


WASH AND OTHER
AREAS, OREG

OTHER STATES 3/


TOTAL


1980-81
1981-82


1980-81
1981-82


: 1980-81
: 1981-82


1980-81
1981-82


2,880


2,220
14,610
17,240
2,355
2,860

33,615
39,040


16.845


2,890


41.325


3.510


19.725


3.745


50.210


4,245


23.710


4.565


60.630


4,980


27,405


5,205


71,160


5,885


31,500


6,105


82,700


6,660


39,020


7,550


105,950


- ------a-a--a----- ------------------------- -sea a a ------ a -- ---- --- a a Ca flea ------- -------- a a as- -


TOTAL QUANTITY RECEIVED AND USED FOR


DOES NOT


PROCESSING REGARDLESS OF THE


INCLUDE QUANTITIES USED FOR POTATO CHIPS


STATE


IN WHICH THE


IN MAINE, MICH OR MINN.


POTATOES WERE
INCLUDES MAINE


GROWN POTATOES ONLY


. 3/


MICH, MINN AND N DAK.


I/


PRODUCED.




TABLE 3.


POTATOES:


PROCESSORS


PRODUCTION AND TOTAL STOCKS OF FALL POTATOES HELD BY GROWERS,


AND LOCAL DEALERS ON DECEMBER


1980 AND DECEMBER


------- flfl~SSS --- a --------------- ------ ------------------------------- ------------------- --- S


CROP OF 1980


CROP OF


STATE


PRODUCTION


* S
:TOTAL STOCKS:


:DEC 1


DEC 1


STOCKS:


AS % OF


: PRODUCTION


: PRODUCTION


:TOTAL STOCKS:


:DEC 1


:DEC 1


STOCKS


AS % OF


: PRODUCTION


,000 CWT


PERCENT


1,000 CWT


PERCENT


CALIF
COLO
IDAHO
MAINE


- LI
- UPSTATE


N DAK


15 STATE TOTAL


6,438
10,950
79,840
24,960
7,403
9,920
1,725
1,876
4,794
6,250
15,680
1,995
19,745
4,180
43,935
16,000
255,691


4,200
7,850
59,000
18,900
4,500
8,400
1,630
1,400
1,850
3,100
10.700


6,734
12,000
80,040
26,520
7,050
13,300
1,739
2,252
5,365
6,875
20,125
1,845
21,710
5,250
52,380
18.190


13,300
3,300
24,500
8.600


171.730


281,375


4,300
8,500
57,500
21,300
5,000
9,100
1,650
1,850
1,860
4,050
14,500
680
15,200
4,000
29,200


187,790


RELIABILITY OF DECEMBER


PRODUCTION AND STOCKS ESTIMATES


To assist


report
below.
final


users


the "Root


e<


deviation


This is computed
estimates as a p
ns for the 196


statistically 1
expected errors
affecting this


"Root


in evaluating


Square


Error"


by expressing


percent
1-80 1


of tt
twenty
Square


in the current estimate


year's


estimates


the reliabi


of production


and stocks


, a statistical measure based on past


the deviations


final


estimate


period;


Error". P
s relative


are not different


"Root Mean Square Error" for the December


chances


are 2


out of 3


that


the final estimate by more than


percent


confidence


the current


1 stocks
estimate


1.9 percent or


level


that


between
s and
square


'robability
to the fin


from those
estimate i


the December
averaging the
root of t


statements
al estimates


of recent


estimates


performance,


estimates


squared


average


can be made
., assuming t


years.


.9 percent


of 188 million cwt will


approximately 3.6 mi
he difference will


in this
s shown
and the


percentage
le becomes
concerning
at factors


r example, tne
means that the


not be above or below


lion cwt.
not exceed


Chances
3.4


approximately 6.4 million cwt.


are 9 out
percent or


Also shown below


production


and stocks


changes between t
3.07 million cwt,


December


is the 10-year


estimates


:he December


1971-80)


and the final


estimate and


record of
estimates


the final


the differences


Using


stocks


estimate during the


ranging from 0.52 million to 6.25 million cwt


estimate has been below the


final


. During ti


estimate 9 out of 10 years.


between the December


again


an example,


10 years have averaged
his 10-year period the


RELIABILITY OF DECEMBER 1


ROOT MEAN
SQUARE ERROR


FALL POTATO ESTIMATES


TEN YEAR RECORD OF DIFFERENCE BETWEEN


DEC 1


AND FINAL ESTIMATE


AND ESTIMATE


: : 90% CONFIDENCE
: : LEVEL
:PERCENT:.....---------...


: PERCENT


QUANTITY


QUANTITY


: NUMBER OF YEARS


: BELOW


: AVERAGE


:SMALLEST:


LARGEST:


FINAL


: ABOVE
: FINAL





TABLE


FALL POTATOES:


STOCKS


BY TYPE


AS PERCENT OF TOTAL STOCKS, DECEMBER


11 MAJOR STATES


1. 1981


- ---------- -------------------


STATE


- -------------------------------------------- ----------------- -----------------------------------------
POTATO TYPES
----------------------- ---- --------------------------------- ...............................................


WHITES


RUSSETS


TOTAL


- ------------------------------ ---------------------------------------------------------------------------------------------------------- -
PERCENT


85
1/2/100
29


COLO
IDAHO
MAINE
MICH 4/
MINN


50
2/3/100
64


11 STATE TOTAL : 4 0 i- w

1/ "INCLUDES SMALL QUANTITIES OF WHITES, LESS THAN 5 PERCENT OF TOTAL.
f/ INCLUDES SMALL QUANTITIES OF REDS, LESS THAN 5 PERCENT OF TOTAL.
3/ INCLUDES SMALL AMOUNT OF RUSSETS, LESS THAN 5 PERCENT OF TOTAL.
1I NOT PUBLISHED TO AVOID DISCLOSURE OF INDIVIDUAL OPERATIONS.


FALL POTATOES


AREA HARVESTED


STATE


YIELD


PRODUCTION


---------- ---- ID S ----- INa--- S-- -- ; -e--ae--a-a -
: IND 2 3 PlOt S E ND


1980


1981 2 1979 : 1980 : 1981


1980


1,000 ACRES


1,000


N
HO SW (10 COUNTIES)
HO OTHER AREAS


MINN
MONT
NEBR
NEV
N Y
N Y
N DAK
OHIO
OREG
OREG
PA
R I
S DAK
UTAH


LONG ISLAND
UPSTATE


6,364
11,455
462
10,050
75,000
1,071
27,685
748
8,000
12,920
1,800
1,482
4,950
6,431
6,463
18,240
2,400
4,610
20,500
6,000
759
1,203
1,375
147
48,450
17,010
1,144
296,919


MALHUER
OTHER AREAS


S 5
S 1,071


,6
87.0
50.0
5.7
981.2


1068
53
5
1,047


6,438
10,950
405
7,850
72,020
S726
24,960


4,180
736
1,072
1,170
120
43,935
16,000
1,340
66,428


6,734
12,000
486
7600
72,240
615
26,52"
743
7,050
13,300
1,739
2,252
3,480
5,365
6,875
20,125
1,845
3,450
18,860
5,250
600
702
1,276
147
52,380
18,190
1,060
90,e664


--------------- ----- ---- ------- ----- ----- ---- ------- ---------- -------- 5 snsefleaseaaaeaaeafle


POTATO STOCKS, DECEMBER 1981


CROP REPORTING BOARD,




TABLE 6.


AREA PLANTED, FALL POTATOES


----------------- ---- a ---------- ------------ a --------------- ---- ee-----
a


STATE


1979


1980


1,000 ACRES


CALIF
COLO
CONN
IDAHO

IND
MAINE
MASS
MICH
MINN


MONT
NEBR
NEV
NY-

N DAK
OHIO
OREG


40.0


- SW CO.
- OTHER CO.


LI
UPSTATE


- MALHEUR CO.
- OTHER CO.


30.0
305.0


116.0
3.5
33.0
70.0
7.5
6.0
15.0
22.0
25.5
121.0
10.4
13.5
52.0
25.0


37.0
1.8
23.0
282.0


108.0
3.4
32.5
65.0
7.0


RI
S DAK
UTAH
VT


WASH :
WIS :
WYO :


.7
102.0
57.0


87.0
52.5


TOTAL


1,100.7


1,002.3


1,070.


1981 POTATO OBJECTIVE YIELD SURVEY


The
the 11
fields


Statistical
major fall
that were


Reporting
producing
selected (


Service conducted potato objective
States in 1981. Sample plots were


a random


basis


sampling procedure. Field workers recorded coui
fields just prior to and immediately after harvest.


using


nts


yield


located


scientifically
measurements f


surveys


n potato
designed


rom


these


data


tables are presented
official estimates.


from
to


the objective
provide current


yield survey
information


presented
about the


in
crop


the
and


following


18.2
40.5
1.8


sample




TABLE 7:


POTATOES:


HARVEST LOSS BY TYPE OF POTATOES,


1980-81 1/


------- ------------------------------------------------------- S S -------------------- ------------- --------- a


STATE


ROUND REDS


ROUND WHITES


RUSSETS


ALL TYPES


------- -------------------- ------------ ----------------------------- ------------------------- --------- ---- -
CWT PER ACRE


IDAHO


MAINE


MINN


N DAK


POTATOES LEFT IN THE FIELD AT TIME OF HARVEST.


INSUFFICIENT SAMPLE SIZE.


TABLE 8:


POTATOES:


AVERAGE NUMBER OF HILLS PER ACRE, BY TYPE,


1980-81


ROUND REDS


ROUND WHITES


AV. NO.
HILLS
PER ACRE


NUMBER
OF
SAMPLES


HILLS
PER ACRE


RUSSETS


NUMBER
OF
SAMPLES


AV. NO.
HILLS
PER ACRE


----------------------------------------------- ------- ------------------------ ------- ----- --------- ---- -
S


12,144
13.057


10,736
9,876

14,309
12,244

10,037
9.246


11,520
11,479

12,577
12,570

10,783
10,370

11,489
10,430

9,910
10,623


16,014
15,018

12,223
12,302

10,119
9,813

13,489
13,970


12,722


13,600
13.709


In C. t


12.63R


STATE :


YEAR


NUMBER
OF
SAMPLES


IDAHO


MAINE


MINN


N DAK




timated


than


tocks


are up
million
ago. I
earlier
from De


stocks


year 4
are up


cwt, 5
Idaho's
. Hold
cember 1


earlier


percent


5 percent,
percent
stocks arn
ings in W
. 1980.


Central


virtually
compared
respective


greater


States


the
with


than


e estimated


ashington


same


are 40
as on


year


y. Holding
on December


at
mnd


Oregon


million
are up


2 million
December


Minn


the
1980
cwt
19


cwt, 1
. 1979


sota


Western


8 percent
percent 1


percent
North D
Wisconsin


States


percent,


below


greater


akota
stoc


tota
two


ks


l 116
years


than a year
respectively,


appearance


cwt, 11
potatoes
grading,


percent
dumped


lost


ito D
above
during


ecemb


)er 1l
same


grading


to shrinkage


period


(moisture


the 15 m
year ago.
livestock
loss).


major
Thi


fall


on pota


States
includes
to farm


total


million
scared


million


without


Potatoe


39.0 million


process
cwt, 16


December


cent more


than


same


seven


period


major


year


ocess


States


total


ago.


FALL


PRODUCTION UP


PERCENT


Pr


1979


a


oduct
3.2 i
Harv


percent
278 cwt


ion


million
ested a


greater


fall p1
metric


ire a


than


acre was


potatoes
tons),


percent


. totaled


1980


percent


cwt above


last


. (24
above


1.05
less


States


million


than


year


1980
acres


1979
above


estimate


crop,


perc


thousand


The a
the 1979


average
yield


million


hectare


. yield


seven


percent


area


years
above


1980


above


total
ago.
1979
crop.


acreage
better
caused


was
than
some


last
178 t


Average


Eastern
year's
thousand


states,


drought
acres,


production


reduced


down


yield


Production


Late
not
last


season r
harvested


year. weti
of acreage


Maine


gains


created


because of
Conditions
and reduce


crop


percent


per acre
placed at


problems


wet cond
in Upst
I yields.


estimated


from


percent
1980 and


million


with


lition


at


million


1979


percent


above
cwt,


Maine


However
York dur


last


Harvested


than


year


percent


potato


, yield


ing


crop1
were


latter


above


some


generally


October


Production


from


eight


1980


percent


Central


above


States
1979.


estimated


average


million


cwt,


acre was


25 cw
294 t
North


t above
thousandd


a year
acres,


cwt greater


percent


more
t at
week


than


than t
1980 I
million


1979


yield


lightly
28 per
d consi
while


Harvested


than
above


twoc
198


area tot
) years
0. Free


al


ago.


cons in


, up


- -. U *-~ *W V~ .- .


a


D--K W 1_


ql ,,


I





UNITED STATES DEPARTMENT OF AGRICULTURE


WASHINGTON, D.C. 20250


POSTAGE
U.S. Ce


UNIVERSITY OF FLORIDA

f3 122 080 487
3 1282 08S60 4870


/FIRS CLAS
FIRST CLASS


OFFICIAL BUSINESS
PENALTY FOR PRIVATE USE, $300


To stop mailing
add res 0


0 or to change your
end this sheet with label


Intact, showing new address, to Crop
Reporting Board Publications, SRS, U.S.


Dept. of Agriculture.


RAm 5829 South


Building. 14th & Independence Ave. S.W.
Wash., DC. 20250


4






~ 4
4


- ,




Full Text

PAGE 1

EURASIPJournalonAppliedSignalProcessing2005:9,1400c 2005HindawiPublishingCorporationDisorderedSpeechAssessmentUsingAutomaticMethodsBasedonQuantitativeMeasuresLingyunGuComputationalNeuroEngineeringLaboratory,DepartmentofElectrical& ComputerEngineering,UniversityofFlorida,Gainesville,FL32611-6200,USAEmail:lygu@cnel.u.eduJohnG.HarrisComputationalNeuroEngineeringLaboratory,DepartmentofElectrical& ComputerEngineering,UniversityofFlorida,Gainesville,FL32611-6200,USAEmail:harris@cnel.u.eduRahulShrivastavDepartmentofCommunicationSciences& Disorders,UniversityofFlorida,Gainesville,FL32611,USAEmail:rahul@csd.u.eduChristineSapienzaDepartmentofCommunicationSciences& Disorders,UniversityofFlorida,Gainesville,FL32611,USAEmail:sapienza@csd.u.eduReceived2November2003;Revised6August2004Speechqualityassessmentmethodsarenecessaryforevaluatinganddocumentingtreatmentoutcomesofpatientssu eringfromdegradedspeechduetoParkinsonsdisease,stroke,orotherdiseaseprocesses.Subjectivemethodsofspeechqualityassessmentaremoreaccurateandmorerobustthanobjectivemethodsbutaretime-consumingandcostly.Weproposeanovelobjectivemeasureofspeechqualityassessmentthatbuildsontraditionalspeechprocessingtechniquessuchasdynamictimewarping(DTWandtheItakura-Saito(IS)distortionmeasure.Initialresultsshowthatourobjectivemeasurecorrelateswellwiththemoreexpensivesubjectivemethods.Keywordsandphrases:objectivespeechqualitymeasures,subjectivespeechqualitymeasures,pathology,anthropomorphic.1.INTRODUCTIONTheaccurateassessmentofspeechqualityisamajorresearchproblemthathasattractedattentionintheeldofspeechcommunicationsformanyyears.Thetwomajorclassesofmethodsemployedintheassessmentofspeechqualityaresubjectiveandobjectivespeechqualitymeasures.Subjectivequalitymeasuresaremoreaccurateandrobustsincetheyaregivenbyprofessionalpersonnelwhohavereceivedspe-cialassessmenttraining,buttheyarenecessarilytimecon-sumingandcostly.Onthecontrary,objectivequalitymea-sures,inspiredbyspeechsignalprocessingtechniques,pro-videane cient,economicalalternativetosubjectivemea-sures.Althoughitisnotsuggestedtouseobjectivequalitymeasurestocompletelyreplacesubjectivemeasures,objec-tivequalitymeasuresdoshowthestrongabilitytopredictsubjectivequalitymeasuresandtheresultsdocorrelateverywellwiththoseproducedbysubjectivequalitymeasures[1 ]. Traditionally,objectivemeasureshavebeenusedtoevaluatespeechafterdecodingandinthepresenceofnoise.Currently,somepioneershavealreadydevelopedsomesystemprotocolsoralgorithmstoapplyobjectivespeechqualityassessmentintodisorderedspeechanalysis.Anymeaningfulqualityassessmentshouldbeconsistentwithhumanresponsesandperception.Therefore,subjectivemeasuresnaturallybecametherstchoicetoevaluatespeechquality.Performancemethodsusingsubjectivemeasuresarebasedonagroupoflistenersopinionofthequalityofanutterance.Subjectivemeasuresusuallyfocusonspeechintel-ligibilityandtheoverallquality.Subjectivemeasurescanalsobebroadlygroupedintotwocategories:utilitarianandana-lytic.Utilitarianmethodshavethreegoals:(1)theyshouldbereasonablye cientintestadministrationanddataanalysis;(2)theyevaluatespeechqualityonaunidimensionalscale;

PAGE 2

DisorderedSpeechEvaluation1401 (3)theymustbereliableandrobustintheirtestmethod.Thekeyaspectofutilitarianapproachesisthattheresultsaresummarizedbyasinglenumber.Ontheotherhand,analyticmethodstrytoidentifytheunderlyingpsychologicalcompo-nentsthatdetermineperceivedquality,andtodiscovertheacousticcorrelatesofthesecomponents.Thereforethere-sultsfromanalyticmethodsaresummarizedonamultidi-mensionalscale[1 ]. ThemodiedrhymetestMRTbyHouseandthedi-agnosticrhymetestDRTbyVoiersarebothintelligibil-itymeasures.Themeanopinionscore(MOS)testandthediagnosticacceptabilitymeasure(DAM)areoverallqualitymeasures,eventhoughMOSisalsocommonlycategorizedasutilitarianandDAMisclassiedasanalytic.Itisunder-standablethatsubjectivequalitymeasuresarethepreferablemeansofqualityassessmentbutsubjectivemeasuresdohaveseveralmajordrawbacks:(1)subjectivemeasuresrequiresig-nicanttimeandpersonnelresources,makingitdi culttoevaluatetherangeofpotentialspeech/voicedistortion;(2)subjectivemeasuresdonotworkverywellwhenthetestedspeechdatabaseislarge[2 ];(3)someratingscoreprotocolsarenotsuitableformeasurementofspeech/voice[3 ];(4)someliteraturesuggeststhatlistenerscannotagreeonspe-cicspeech/voiceratings[4 ]. Comparedwiththesubjectivemeasuresmentionedabove,objectivemeasureshaveseveraloutstandingadvan-tages:(1)theyarelessexpensivetoadminister,savingmoney,time,andhumanresources;2)theyproducemoreconsis-tentresultsandarenota ectedbyhumanerror;(3)mostimportantly,theformoftheobjectivemeasureitselfcangivevaluableinsightintothenatureofthehumanspeechper-ceptionprocess,helpingresearchersunderstandthespeechproductionmechanismmoredeeply[1 ].Generallyspeaking,objectivespeechqualitymeasuresareusuallyevaluatedinthetime,spectral,orcepstraldomains.Thispaperisorganizedasfollows.InSection2,dis-orderedspeechbackgroundwillbeintroduced.Then,inSection3,theDTWmethodisdiscussed.Specicspeechfea-turesfordisorderedspeechwillbeproposedinSection4. Section5dealswithonesubjectivemeasure.Allexperimen-talresultsarediscussedinSection6.Finally,conclusionsaredrawninSection7. 2.DISORDEREDSPEECHBACKGROUNDUsually,patientswithParkinsonsdiseaseorpeoplewhohavesu eredastrokehavedi cultyproducingclearspeech,re-sultinginalossofintelligibility.Hence,itisimportanttodevelopameanstohelpthemproducemoreclearspeechordevelopalgorithmstoautomaticallyclarifytheirunclearspeech.Thesee ortsrequireane cientmethodtoevaluatedisorderedspeechastherststep.Attemptstodevelopalgorithmstoevaluatedisorderedspeechrequireustounderstandhowdisorderedspeechisproduced,thefactorsthata ectdisorderedspeech,andtheexplicitphenomenarelatedtothesefactors.Thetermdysarthriaisusedtodescribechangesinspeechproduction Automaticassessment procedurePatientsspeechDTW alignmentObjectivequalityassessment HealthyspeechScorescaling systemSpeechqualityscore Figure1:Objectivepatientsspeechqualityassessmentblockdia-gram.characterizedbyanimpairmentinoneormoreofthesys-temsinvolvedinspeech[5 ].Thethreemajorsystemsin-volvedinspeechproductionarerespiration,voiceproduc-tion,andarticulation.Voiceisproducedbythelarynxandtheoralstructuresarticulatetomodifythesoundsourcepro-ducedbythelarynx.ThedysarthriaassociatedwithParkin-sonsdiseaseisreferredtoasahypokineticdysarthria[6 7 ]. Commonsymptomsofhypokineticdysarthriaincludere-ducedloudnessofspeechand/ormonoloudness(lackofloudnessvariation)andreducedspeakingratewithintermit-tentrapidburstsofspeech.Forinstance,speakersmayshowaslowrateofspeech,butparticularwordsorphraseswithinthatutterancemaybeproducedwitharapidrate.Theoralstructuressuchasthetongueandlipsarerigid,resultinginareducedrangeofmovement.Thise ectivelydampensthespeechsignalanddistortstheaccuracyofthesoundcon-sonantorvowel)production.Theremaybesomeinstancesofhypernasalityastheconditionworsensresultingfromaninadequatevelarclosure.Thismayalsoresultinthedamp-eningofthesoundproduced.Voicequalityinthesepatientsisoftendescribedashoarseorharsh.Inthispaper,wetestseveralwell-knownspeechprocess-ingparametersthatcanquantifytheseverityofdisorderedspeech.ThesearetheItakura-SaitoIS)measure,thelog-likelihoodratioLLR)measure,andthelog-area-ratio(LAR)measurewhichevaluatethespectralenvelopeofthegivendisorderedspeech.Figure1showstheobjectivedisorderedspeechqualityassessmentblockdiagram.3.DYNAMICTIMEWARPINGConventionalobjectivespeechqualitymeasuresareusedtoevaluatethespeechqualityafterspeechiscodedanddecodedortransmittedwithnoiseandchanneldegradation.Inthesescenarios,theoriginalhigh-qualityspeechandthedegradedspeechhaveexactlythesamelength,whichleadstoasimpleone-to-onecomparisonofwindowsfromeachspeechutter-ance.However,inthisproject,weusethespeechproduced

PAGE 3

1402EURASIPJournalonAppliedSignalProcessing byhealthypeopleasthegoldstandardtocomparewithdis-orderedspeech.Inthiscase,aligningthetwodi erentspeechsegmentstothesamereasonablecomparablelengthiscru-cial.Dynamictimewarping(DTWisthemoststraightfor-wardsolutionandisusedtosolveexactlythisprobleminspeechrecognitionapplications.Giventwospeechpatterns,X and Y ,thesepatternscanberepresentedbyasequencex 1 x 2 ... x T x )andy 1 y 2 ... y T y ),wherex i and y i arethefeaturevectors.Aswehavenoted,ingeneralthesequenceofx i swillnothavethesamelengthasthesequenceofy i s.Inordertodeterminethedistancebe-tweenX and Y ,giventhatsomedistancefunctiond x y )ex-ists,weneedameaningfulwaytodeterminehowtoproperlyalignthevectorsforthecomparison.DTWisonewaythatsuchanalignmentcanbemade[8 ].Wedenetwowarpingfunctions, x and y ,whichtransformtheindicesofthevec-torsequencestoanormalizedtimeaxis,k .Thuswehavei x = x k ), k = 1,2,... T i y = y k ), k = 1,2,... T. (1) Thisgivesusamappingfromx 1 x 2 ... x T x )to(x 1 x 2 ... x T )andfromy 1 y 2 ... y T y )to(y 1 y 2 ... y T ).Withsuchamapping,weareabletocomputed x y usingthesewarp-ingfunctions,givingusthetotaldistancebetweentwopat-ternsasd x y = T k = 1 d x k ), y k m k M ,(2wherem k isapathweightandM isanormalizationfactor.Thus,allthatremainsisthespecicationofthepath indicatedintheaboveequation.Themostcommontechniqueistospecifythat istheminimumofallpossiblepaths,subjecttocertainconstraintsbyusingtheequationasfollows:d X Y min d x y (3) Fortimenormalization,theoptimalpathbasedonDTWhasedbeginningandendingpoints.Someotherconstraintsmayalsoapply.Forexample,thepathshouldbemonotonic,whichrequiresapositiveslope.Thisconstrainteliminatesthepossibilityofreversewarping.Therefore,wechoosetoen-forcetheTypeIIIlocalconstraint[8 ].Inaddition,ournu-merousexperimentalresultsshowthattheeightlocalcon-straintswillnotsignicantlychangethenalresults.Becauseofthelocalcontinuityconstraints,certainportionsareex-cludedfromtheregiontheoptimalwarpingpathcantra-verse.Byusingthemaximumandminimumpossiblepathexpansion,wecandeneglobalpathconstraintsasfollows:1+ x k 1 Q max 1+ Q max x k 1 T y + Q max x k T x T y + x k T x Q max (4) Inthisaspect,slopeweightingalongthepathaddsyetan-otherdimensionofcontrolinthesearchfortheoptimalwarpingpath.Therearefourtypesofslopeweighting.Thetypechoseninthispaperism k = x k x k 1)+ y k y k 1) (5) Ifwetakethenotationd i x i y asthedistancebetweenx ix and y iy ,whicharetheelementsofx 1 x 2 ... x T x )and y 1 y 2 ... y T y ),respectively,andD i x i y astheaccumula-tiveoptimalvalue,thenwecanapplytheexactlocalcon-straintaswellastheslopeweighttogetD i x i y = min D i x 2, y x 1 +3 d i x i y D i x 1, y x 1 +2 d i x i y D i x 1, y x 2 +3 d i x i y (6) 4.OBJECTIVEQUALITYMEASURESFromananthropomorphicperspective,speechproductionisverycomplexbutasimpleviewisthatvowelsareproducedbythelungs,thelarynxexcitation,andtheresonanceofthevocaltract.Thelaryngealcongurationandthetonguespo-sitiondramaticallychangeanindividualspeakesspeechin-tonation,pitch,orquality.Forexample,duetodi erencesintonguepositionsduringpronunciation,nonnativespeakersofEnglishmayusetonguemovementscharacteristictotheirnativelanguage,therebyproducinganoticeableaccent.Sim-ilarly,therigidtonguemovementoftheParkinsonspatientcausestheirpronunciationtobecomedistorted.Weattempttodevelopobjectivespeechqualitymeasuresusingknowl-edgeofhumanspeechproduction.However,werstneedtodeneafewtermscommonlyusedinspeechprocessing.Aformantisdenedasapeakinthespeechpowerspectrum.Thepitchofspeechisusuallydeterminedbythefrequencyoftheexcitationsignal,whichisproducedbythevibrationofthevocalfolds.Thevocaltractresonanceisusuallyrepre-sentedbythespectralenvelope.Somecontemporaryresearchhasalreadymadeprogressonobjectiveanalysesofdisorderedspeech.Forinstance,theComputerizedSpeechLab(CSL)producedbyKayElemet-ricsCorporationisacommerciallyavailablehardwareandsoftwarepackagefortheanalysisofdisorderedspeech.TheCSLallowsacliniciantocalculateseveralmeasuresrelatedtotheintelligibilityandqualityofdisorderedspeech.An-othercommercialproductistheEVAsystem,madebySQ-Lab,Marseille,France.Thissystemallowssimultaneousmea-surementofacousticandaerodynamicparametersrelatedtospeechproduction.Acousticsignalsarerecordedusingthemicrophonebuiltintothepneumotachographwhichisusedtomeasureoralairow.Intraoralpressuremaybecalculatedusingabuilt-inpressuresensor[9 ].Themajorityofsuchanalysispackagesallowthecalculationofacousticandaero-dynamicparameterssuchasjitter,shimmer,signal-to-noiseratio,oralairow,andvoiceonsettime.However,thecon-cordancebetweentheseobjectivemeasuresandperceptualratingsofqualityandintelligibilityremainsatarelativelylow

PAGE 4

DisorderedSpeechEvaluation1403 percentage[10 11 ],andisoftenunsuitableforclinicalpur-poses.Manyofthesemeasurescanonlybecalculatedfromrelativelysteadyportionsofthespeechsignal.However,nu-merousstudieshavestressedthattheunsteadypartsofthesignal,suchasonset,couldprovidevaluableinformationforobjectiveevaluationofspeechandallownerdiscriminationoftheseverityofdysphonia.Inaddition,manyofthesemea-suresarecalculatedfromasinglevowelthatpatientsarere-quiredtoproduceforarelativelylongperiodoftime[12 13 ]. Inreality,thenaturalcontinuoussentencemayprovideamoreaccuratepictureofthepatientsspeechdisorder.Toovercomesomeoftheseshortcomingsoftheexistingspeechanalysistechniques,weproposeanewalgorithmorig-inallyinspiredbythespeechcoding-decodingandspeechtelecommunicationstechniques.Therstmeaningfulmea-surewhichcanbeobtainedtocomparespeechdi erencesistocomputethedi erencesofthelogarithmsofthepowerspectrumateachfrequencyrange[4 ].Weusethefollowingequationtorepresentthedi erence:d w = ln X w 2 ln Y w 2 ,(7whereX w )andY w arethemagnitudesinthefrequencydomainoftwocomparedspeechsignals.Itisalsopossibletoformallyexpressthemosteasyandstraightforwardmethodtostandforthespectraldistortionasfollows:d X Y = d w k dw 2 1 /k ,(8where,again,X and Y hererepresentthetwospeechsignalstobecompared.Althoughtheabovemethodiseasytoimplement,goodresultsarenotguaranteed.Manydi erenttypesofmodi-standardobjectivequalitymeasureshavebeenproposed.TheseincludemeasuressuchastheItakura-SaitoIS)dis-tortionmeasure,thelog-likelihoodratioLLR)measure,thelog-area-ratio(LAR)measure,thesegmentalSNRmeasure,andtheweightedspectralslopeWSS)measure.Inthispa-per,wechosetoinvestigatetherstthreemeasures:IS,LLRandLAR[14 15 16 ]. TheISdistortionmeasureiscalculatedbasedonthefol-lowingequation:d IS a d a = 2 2 d a d R a T d a R a T +log 2 2 d 1,(9)where 2 and 2 d representtheall-polegainsforthestan-dardhealthypeoplesspeechandthetestpatientsspeech.a and a d arethehealthy-speechandpatient-speechLPCcoef-cientvectors,respectively.R istheautocorrelationmatrixfor x n ),wherex n isthesampledspeechofhealthypeo-ple.TheelementsofR aredenedasr | i j | = N i j | n = 1 r n r n + | i j | | i j |= 0,1,... p (10) Table1:MOSsubjectivemeasureevaluationtable. RatingSpeechqualityLevelofdistortion 5ExcellentImperceptible4GoodPerceptible,butnotannoying3FairPerceptible,andslightlyannoying2PoorAnnoying,butnotobjectionable1UnsatisedVeryannoyingandobjectionable whereN isthelengthofthespeechframeandp istheorderofLPCcoe cients. LLRissimilartotheISmeasure.However,whiletheISmeasureincorporatesthegainfactorbyusingvarianceterms,LLRonlyconsidersthedi erencebetweenthegeneralspec-tralshapes.ThefollowingequationprovidesthedetailsforcomputingtheLLR:d LLR a d a = log a d R a T d a R a T (11) LARisanotherspeechqualityassessmentmeasurebasedonthedissimilarityofLPCcoe cientsbetweenhealthyspeechandthepatientsspeech.Di erentfromLLR,LARusestheretioncoe cientstocalculatethedi erenceandisexpressedbytheequationd LAR = 1 p p i = 1 log 1+ r i 1 r i log 1+ r d i 1 r d i 2 1 / 2 ,12)wherep istheorderoftheLPCcoe cients, r i )andr d i )arethe i thretioncoe cientsofhealthyandpatientsspeechsignals.Inthefollowingsectiondescribingtheexperimentandresults,wewillcomparetheperformancesofeachofthesemeasuresappliedtoourdatabase.Thecorrelationbetweentheseobjectivequalityassessmentmeasuresandonesubjec-tivequalityassessmentwillalsobediscussed.5.SUBJECTIVEQUALITYMEASURESNomatterhowspeechqualityisdened,itmustbebasedonhumanresponseandperception.Sodesigningasuitablesubjectivemeasureofqualityisveryimportantintheassess-mentofspeechquality.Correspondingly,themostimportantcriteriontoevaluatetheaccuracyofanobjectivemeasureofqualityistodetermineitscorrelationwithsubjectivequalitymeasures.AsdiscussedinSection1,subjectivemeasurescanbebroadlydividedintoutilitarianandanalyticcategories.Withoutlossofgeneralization,wewillusetwooftheutili-tarianmethodsforourinvestigation.Onereliableandeas-ilyimplementedsubjectiveutilitarianmeasureisthemeanopinionscore(MOS)[1 4 ].Inthismethod,humanlisten-ersratethespeechundertestontheve-pointscaleshownin Table1.Relatedresearchshowsthatasfewasvebutnomorethanninecategoriesareenoughfortheassessmentof

PAGE 5

1404EURASIPJournalonAppliedSignalProcessing Table2:Moderate-severesubjectivemeasureevaluationtable. RatingLevelofdistortion 3Moderate2Moderatetosevere1Severe quality.Thenalspeechqualityassessmentvaluecanbecal-culatedastheaverageoftheresponsesofseverallisteners.TheMOStestiswidelyusedinthetelecommunicationsareatocomparetheoriginalsignalqualitywiththatofthedis-tortedsignal.Fordisorderedspeechanalysis,however,itmaynotbefeasibletocategorizesentencesasperceptible,butnotannoyingorannoying,butnotobjectionable.Therefore,adi erentcommonlyusedsubjectiveutilitarianmeasurewasobtained.Inthistest,listenersratedthesentencesintothreecategories:mild,moderate,orsevere[5 6 7 ].Asimilar4-pointratingscale,calledtheGRBASmethod,hasbeenpre-sentedfortheevaluationofdisordervoicequality[17 ].Inthesesubjectivetests,eachtestsentencewasassignedascorebasedonwhetherthedisorderedsentencequalitywasper-ceivedtobemild,moderate,orsevere.BasedonourdatabaseofParkinsonspatientstestedinthisexperiment,wemodi-themild-moderate-severeratingscaletohavethreenewlevels:moderate,moderatetosevere,andsevere.ThedetailsandcriteriafortheseratingsarelistedinTable2.Thefol-lowingprocedureswerefollowedwhenobtainingperceptualjudgmentinthepresentexperiment:Listenerswereaskedtolistencarefullytoeachtestsentence.Listenerswereallowedtohearthetestsentenceasmanytimesasneededtoensurethattheyassignedthemostappropriatescoretoeachsentence.ListenerswereaskedtoreadthecriteriatableTables1 and 2 carefullyandwererequiredtoassignascoretoeachsen-tencebasedonthelevelofdistortiondescribedinthetables.6.EXPERIMENTALRESULTSThespeechdatabaseusedinthisexperimentwascollectedbytheexperimentersattheMotorMovementDisordersClinic,UniversityofFlorida.TenpatientswithParkinsonsdiseasewererecordedreadingastandardpassageGrandfatherPas-sageAdditionally,thesamepassagewasalsorecordedfromfourhealthyadultspeakers.Althoughspeakersvaryintheirrateofspeech,thispassagetakesapproximately1minutetoread.Threesuccessivesentencesaround15secondsindu-ration)wereselectedfromthispassageforacousticandper-ceptualanalyses.ThesentencesincludeYouwishtoknowallaboutmygrandfather.Well,heisnearlyninetythreeyearsold.Hedresseshimselfinanancientblackfrockcoat,usu-allyminusseveralbuttons.Thefourteenspeakersweredi-videdintotwogroupsmalesandfemales.Intherstlis-teningtest,sixlistenersevaluatedthespeechoffourParkin-sonspatientsandonehealthyspeaker.Inthesecondlisteningtest,wetestedtwelvelistenerswhoratedthespeechofsevenParkinsonspatientsandonehealthyspeaker.Ofthe18par-ticipantsinthelisteningtests,sixwerefromtheUSA,vefromChina,vefromIndia,onefromKorea,andonefromTurkey.Sevenofthemweremaleandtherestwerefemale.AlllistenersspokeuentEnglish.TherstlisteningtestwasusedtoobtainratingsusingtheMOScriterialistedinTable1.Listenersgaveanindivid-ualscoretoeachsentence.Inthisstudy,twodi erentmeth-odswereusedtocomparetheobjectiveandthesubjectivemeasures.Intherstmethod,allMOSscoresgivenbythelistenerswerecorrelatedwiththedistancemeasurescalcu-latedbythevariousalgorithms.Inthesecondapproach,theorderoftheMOSscoresratherthantheactualvalueoftheMOSscores)wascorrelatedwiththedistancemeasures.Inthisapproach,listenerssimplyorderedeachsentencefromthebesttotheworstquality.Iftwoormoresentencesweregiventhesamerank,listenerswereaskedtolistencarefullyandchoosedi erentranksforeachsentence.Incontrast,intherstmethod,listenersmayendupgivingidenticalintegerscorestotwospeechsegmentseventhoughonemaysoundnoticeablybetterthantheother.Table3givesthedetailsonallthesentencesscoredusingMOSscaleformalespeakersonly.SentenceslabelledasP1,P2,P3,andP4werespokenbytheParkinsonspatientsandH1isthesentencesspokenbythehealthyspeaker.ThesixlistenersarelabelledasList1toList6. Onesentencefromahealthyspeakerwasusedasthestandardsentenceforcalculatingtheobjectivemeasuresofquality.DTWwasrstappliedtoalignthisstandardsen-tencewitheachpatientssentence.Figure2showstheopti-malframematchpathbetweenthestandardhealthyspeechandthepatientsspeech.Forthesecondmethod,everypro-cedureisthesameexceptforreplacingtheexactscorebytherelativeorder.Therefore,inTable3,eachcolumnistheordergivenbyeachlistener.Finally,thethreedistortionmeasuresIS,LLR,andLAR)werecalculated.ThelastthreecolumnsinTable3showtheexactvaluesofIS,LLR,andLAR,respectively.InTable4, thelastthreecolumnsshowtherelativeorderofthedistor-tionscoresobtainedfromeachspeaker.Figure3showsthehealthyspeechwaveformupperpanel),thepatientspeechwaveform(middlepanel)andtheirdistortioncurvecalcu-latedbytheISmeasurelowerpanel).Figure4showsasim-ilarcomparisonbasedonLLRandFigure5showsthesamecomparisonbasedonLAR.Figure6exhibitsthehistogramofthedistortionvalues,whichmaygiveusdeeperinsightaboutthedi erencesbetweenthehealthyspeakerandthepatientsspeech.Thismayprovidegreaterinformationthantheuseofasinglenumberobtainedbyaveragingthedistortionmea-suresacrossanumberofframes.Asdiscussedearlier,thequalityofanobjectivemeasureisdeterminedbyhowwellitpredictsthesubjectivemeasure.Thefollowingformulaiswidelyusedtoevaluatetheperfor-manceofobjectivemeasures: = d S d S d O d O d d S d S d 2 d O d O d 2 1 / 2 ,13)whereS d and O d aresubjectiveandobjectiveresults. S d and O d aretheircorrespondingaveragevalues.Table3showsall

PAGE 6

DisorderedSpeechEvaluation1405 Table3:Subjectivetestresultsandtheircorrelationwithobjectivetestusingmethod1intherstround. SubjectList1List2List3List4List5List6Avg.ISLLRLAR P12322322.3371035197.51441.5P22121211.50769990175.61054.2P33111221.67572200152.31014.9P43232332.67304150218.81025.4H155555552415596.2752.5 Corr.0.76380.64190.5729 20 40 60 80 100 120 140 160 180 50100150FramenumberFramenumber 20 40 60 80 100 120 140 160 180 50100150FramenumberFramenumber Figure2:Dynamictimewarping(DTWoptimalpathbetweentherecordedspeechofahealthypersonhorizontalaxis)andaParkinsonspatientverticalaxis).Table4:Subjectivetestresultsandtheircorrelationwithobjectivetestusingmethod2intherstround. SubjectList1List2List3List4List5List6Avg.ISLLRLAR P15233333.2245P24445454.3534P32554544.2422P43322222.3353H11111111111 Corr.0.86840.18280.5142 threeobjectivemeasuresandtheircorrelationvaluesbasedonmethod1.TheISmeasure,withacorrelationof0.7638,showedthebestperformance.Table4liststhecorrelationvaluesbasedonmethod2,andonceagaintheISmeasureshowedthehighestcorrelationof0.8684.Inanalyzing9 ), 11 )and12 ),wecanseethatthegoodperformanceoftheISmeasuremightbepartiallyduetothefactthatitnotonlyconsidersthegeneralspectraldi erence,butalsousesthevariancetermtotakeintoaccountthegainfactoroftheall-poleermodel.Aftercompletingthepreliminarytest,asecondtestwasconductedtovalidateourconclusionthatISisagoodmea-sureofdisorderedspeechquality.Inthistest,speechsamplesfromalargernumberofpatientswithParkinsonsdisease(seveninsteadoffour)wereratedbymorelistenerstwelveinsteadofsix).InadditiontotheMOSscores,listenerswerealsoaskedtocategorizethespeechsamplesasNormal,mod-erate,moderatetosevere,orsevere.TohighlightthevalidityoftheISmeasures,onlythismeasurewascalculatedforthespeechsamplesusedinthesecondtest.Table5showsthe

PAGE 7

1406EURASIPJournalonAppliedSignalProcessing 0 5 0 0 5 1 Amplitude00 511 52 Times 0 5 0 0 5 Amplitude00 20 40 60 811. 21 41 61 82 Times 5000 4000 3000 2000 1000 Amplitude20406080100120140160180200Framenumber Figure3:ISvaluelower)versushealthyspeechwaveformupper)andpatientspeechwaveformmiddle). 0 5 0 0 5 1 Amplitude00 511 52 Times 0 5 0 0 5 Amplitude00 20 40 60 811. 21 41 61 82 Times 15 10 5 Amplitude20406080100120140160180200Framenumber Figure4:LLRvaluelower)versushealthyspeechwaveformup-per)andpatientspeechwaveformmiddle).MOSfromindividuallisteners,theaverageMOS,andthecorrelationbetweentheISmeasureandMOSvaluesbasedonmethod1describedearlier.Thiscorrelationwasfoundtobe0.8032andiscomparablewith0.7638obtainedintherstroundtest.Table6showsthemoderate-severetestscoresfromeachlistener,theaveragemoderate-severetestscores,andthecorrelationbetweentheISmeasureandthesubjec-tiveratings.Onceagain,acorrelationof0.7417wasobtainedwhichiscomparabletothatobtainedintherstroundtest. 0 5 0 0 5 1 Amplitude00 511 52 Times 0 5 0 0 5 Amplitude00 20 40 60 811. 21 41 61 82 Times 2 5 2 1 5 1 0 5 Amplitude20406080100120140160180200Framenumber Figure5:LARvalue(lower)versushealthyspeechwaveformup-per)andpatientspeechwaveformmiddle). 80 70 60 50 40 30 20 10 0 00 511. 522. 53 10 4 DistortionvalueCount Figure6:ThehistogramofthedistortionvaluesbasedontheISmethod. AllobjectivespeechqualityassessmentcriteriaIS,LLR,LARetc.)proposedabovemainlyfocusonthespeechspec-tralenvelope.Fromtheperceptualpointofview,wearemainlyinterestedinhowtoe cientlyevaluatethespeechintelligibilityandquality.However,intelligibilityandqual-ityarenottheonlyaspectsoftheoverallspeechqualityevaluation.Manyotherfactorsthata ectspeechqualitymayalsoneedtobeconsidered.Forinstance,HansenandNand-kumarproposedthatpitchturbulence(PTmaybeusedtoevaluatethemonotoneorpitchvariation,whichisdirectlyrelatedtothelaryngealexcitationsignal.Similarly,energyturbulence(ETisanotherimportantfactorusedtoevaluate

PAGE 8

DisorderedSpeechEvaluation1407 Table5:Subjectivetestresultsandtheircorrelationwithobjectivetestusingmethod1inthesecondroundbasedonMOStest. SubjectList1List2List3List4List5List6List7List8List9List10List11List12Avg.IS P13243223333112.5041500P22332312223212.1784200P31221212112111.42264000P44444435454444.0810300P54433345445443.9229800P61321223123211.92205000P72322234333322.67103000H15555535555554.836010 Corr.0.8032 Table6:Subjectivetestresultsandtheircorrelationwithobjectivetestusingmethod1inthesecondroundbasedonmoderate-severetest. SubjectList1List2List3List4List5List6List7List8List9List10List11List12Avg.IS P11221212111211.42205000P22212222232222103000P3333333333333310300P41111111211111.08264000P52221222122111.6784200P62312223222222.0841500P7333333333333329800 Corr.0.7417 themonoloudnessorenergyvariation[2 ].Thefollowingequationsgivetheexactmathematicexpressionsforthesemeasures:PT = 1 N 1 N 1 i = 1 P i +1 P i ET = 1 N 1 N 1 i = 1 E i +1 E i (14) whereN isthetotalnumberofframesofthegivensen-tence,andP i )andE i representthepitchandenergyoftheframei .Weusedthedataobtainedintherstroundofevaluationtotestthecorrelationbetweenthesemeasures(PTandETandthesubjectiveratings.Table7showsthepitchturbulence(PTandenergyturbulenceETvaluescalcu-latedfrom14 aswellastheircorrelationbasedonmethod1. Table8showsthesimilarresultsbasedonthemethod2.Figures7 and 8 showthepitchturbulenceandenergytur-bulencefromagivenspeechsignal.BasedonTables7 and 8 ,itappearsthatPTandETarepoorlycorrelatedwiththesubjectiveassessments,usingeithermethod1ormethod2.Thissuggeststhatduringsubjectiveassessment,humansputmostoftheiremphasisonintelligibility,which,fromasignalprocessingview,isrelatedprimarilytothespectralenvelope.Theexcitationpitchandenergyvariationarenotasimpor-tantasspectralenvelopevariationintheperceptionofover-allspeechquality.Eveninourcurrentalgorithm,pitchandenergyturbulencewerenotverye cientinpredictingtheTable7:SubjectivetestresultsandtheircorrelationwithPTandETtestusingmethod1. SubjectAvg.PTET P12.3313.27773.9040P21.508.07128.0712P31.674.577511.9782P42.6716.36075.8966H154.28158.3446 Corr.0.12640.1137 Table8:SubjectivetestresultsandtheircorrelationwithPTandETtestusingmethod2. SubjectAvg.PTET P13.245P24.333P34.221P42.354H1112 Corr.0.18280.0800 overallspeechquality.Potentially,eventhoughthecorrela-tionperformancewithone-dimensionalevaluation(suchasMOS)ispoor,thesetwoparametersmaycorrelatewellwithmultidimensionalevaluation(suchasDAM).

PAGE 9

1408EURASIPJournalonAppliedSignalProcessing 0 5 0 0 5 1 Amplitude00 511. 52 Times 200 150 100 Pitch20406080100120140160180200220Framenumber Figure7:Thepitchturbulence(lower)fromagivenspeechsignal(upper). 0 5 0 0 5 1 Amplitude00 511. 52 Times 100 50 0 50 100 150 Amplitude20406080100120140160180200220Framenumber Figure8:Theenergyturbulencelower)fromagivenspeechsignal(upper). 7.CONCLUSIONObjectiveevaluationofdisorderedspeechqualityisnotaneasytask.Inthispaper,wediscussthreeobjectivequal-ityassessmentmeasuresandonesubjectivemeasure.Byevaluatingourspeechdatabase,theISmeasureshowedastrongcorrelationwiththeMOStests.Therefore,theISmea-sureissuggestedtobemoresuitablethanLLRandLARforuseasareliabletooltoevaluatetheoverallqualityofdisor-deredspeech.TheISmeasurecouldalsobeusedtopredictthesubjectivequalitymeasureMOSscoregivenbyhumans.ACKNOWLEDGMENTTheauthorsaregratefultothreereviewerswhoprovideduswithalargenumberofdetailedsuggestionsforimprovingthesubmittedmanuscript.REFERENCES [1]S.Quanckenbush,T.Barnwell,andM.Clements,ObjectiveMeasuresofSpeechQuality,PrenticeHall,NewYork,NY,USA,1988. [2]J.HansenandS.Nandkumar,ObjectivequalityassessmentandtheRPE-LTPvocoderindi erentnoiseandlanguagecon-ditions,JournaloftheAcousticalSocietyofAmerica,vol.97,no.1,pp.609,1995.[3]J.HansenandL.Arslan,Robustfeature-estimationandob-jectivequalityassessmentfornoisyspeechrecognitionusingthecreditcardcorpus,IEEETrans.SpeechAudioProcessing, vol.3,no.3,pp.169,1995.[4]S.Dimolitsas,Objectivespeechdistortionmeasuresandtheirrelevancetospeechqualityassessments,IEEProceed-ingsPartI:Communications,SpeechandVision,vol.136,no.5,pp.317,1989.[5]L.Ramig,C.Bonitati,J.Lemke,andY.Horii,Voicetreat-mentforpatientswithParkinsondisease:Developmentofanapproachandpreliminarye cacydata,JournalofMedicalSpeech-LanguagePathology,vol.2,no.3,pp.191,1994.[6]S.Countryman,L.Ramig,andA.Pawlas,SpeechandvoicedecitsinParkinsonianPlussyndromes:Cantheybetreated?JournalofMedicalSpeech-LanguagePathology,vol.2,no.3,pp.211,1994.[7]S.CountrymanandL.Ramig, ectsofintensivevoicether-apyonvoicedecitsassociatedwithbilateralthalamotomyinParkinsondisease:Acasestudy,JournalofMedicalSpeech-LanguagePathology,vol.1,no.4,pp.233,1993.[8]L.RabinerandB.Juang,FundamentalofSpeechRecognition, PrenticeHall,NewYork,NY,USA,1984.[9]P.Yu,M.Ouaknine,J.Revis,andA.Giovanni,Objectivevoiceanalysisfordysphonicpatients:amultiparametricpro-tocolincludingacousticandaerodynamicmeasurements,JournalofVoice,vol.15,no.4,pp.529,2001.[10]A.Giovanni,D.Robert,N.Estublier,B.Teston,M.Zanaret,andM.Cannoni,Objectiveevaluationofdysphonia:prelim-inaryresultsofadeviceallowingsimultaneousacousticandaerodynamicmeasurements,FoliaPhoniatrLogop,vol.48,no.4,pp.175,1996.[11]J.Revis,A.Giovanni,F.Wuyts,andJ.Triglia,Comparisonofdi erentvoicesamplesforperceptualanalysis,FoliaPhoniatrLogop,vol.51,no.3,pp.108,1999.[12]D.Berry,K.Verdolini,D.Montequin,M.Hess,R.Chan,andI.Titze,Aquantitativeoutput-costratioinvoiceproduction,JournalofSpeech,LanguageandHearingResearch,vol.44,no.1,pp.29,2001.[13]P.Dejonckere,C.Obbens,G.DeMoor,andG.Wieneke,er-ceptualevaluationofdysphonia:Reliabilityandrelevance,FoliaPhoniatrLogop,vol.45,no.2,pp.76,1993.[14]E.WallenandJ.Hansen,Ascreeningtestforspeechpathol-ogyassessmentusingobjectivequalitymeasures,inProc.4thInternationalConferenceonSpokenLanguageProceedingsIC-SLP96),vol.2,pp.776,Philadelphia,Pa,USA,October1996. [15]L.ThorpeandW.Yang,erformanceofcurrentperceptualobjectivespeechqualitymeasures,inProc.IEEEWorkshoponSpeechCodingProceedings(SCW,pp.144,Porvoo,Finland,June1999.[16]J.HansenandB.Pellom,Ane ectivequalityevaluationpro-tocolforspeechenhancementalgorithms,On-lineTechnicalReport.[17]M.Hirano,Psycho-AcousticEvaluationofVoice:GRBASScaleforEvaluatingtheHoarseVoice,SpringerVerlag,NewYork,NY,USA,1981.

PAGE 10

DisorderedSpeechEvaluation1409 LingyunGureceivedhisB.S.andM.S.de-greesinelectricalengineeringfromtheUni-versityofElectronicScienceandTechnol-ogyofChina(UESTCandOldDomin-ionUniversityin1998and2002,respec-tively.HeiscurrentlypursuinghisPh.D.intheComputationalNeuroEngineeringLab-oratory,ElectricalandComputerEngineer-ingDepartment,UniversityofFlorida(UF).Hismainresearchinterestsareinrobustspeechrecognition,speechsignalprocessing,andauditoryproduc-tionandperception. JohnG.HarrisreceivedhisB.S.andM.S.degreesinelectricalengineeringfromMITin1983and1986.HeearnedhisPh.D.de-greefromCaltechintheinterdisciplinaryComputationandNeuralSystemsProgramin1991.Afteratwo-yearpostdocattheMITAILab,Dr.HarrisjoinedtheElectricalandComputerEngineeringDepartment,Uni-versityofFlorida(UF).HeiscurrentlyanAssociateProfessorandleadstheHybridSignalProcessingGroupinresearchingbiologicallyinspiredcir-cuits,architectures,andalgorithmsforsignalprocessing.Dr.Harrishaspublishedover100researchpapersandpatentsinthisarea.HecodirectstheComputationalNeuroEngineeringLaboratoryandhasajointappointmentintheBiomedicalEngineeringDepart-mentatUF. RahulShrivastavearnedhisB.S.degreein1995andM.S.degreein1997inspeechandhearingsciencesfromtheUniversityofMysore,India.HecompletedhisPh.D.degreeinspeechandhearingsciencefromIndianaUniversity,Bloomington,in2001.CurrentlyheisonthefacultyattheDepart-mentofCommunicationSciencesandDis-orders,UniversityofFlorida.Hisresearchisstudyingthefactorsthata ectthepercep-tionofvoicequalityandspeechintelligibilityinpatientswithava-rietyofspeechdisorders. ChristineSapienzareceivedherPh.D.degreeinspeechsciencefromTheStateUniversityofNewYorkatBu aloin1993.Currently,sheisaProfessorintheDepart-mentofCommunicationSciencesandDis-orders,UniversityofFlorida.Hermostre-centworkhasfocusedontheuseofstrengthtrainingparadigmsinmultiplepopulationsincludingvoicedisorders,Parkinsonsdis-ease,spinalcordinjury,andmultiplescle-rosis.Shemaintainsanactiveresearchlaboratorywith7currentPh.D.students.HerclinicalworktakesplaceatAyersOutpatientVoiceClinicandtheMotorMovementDisordersClinicattheUni-versityofFlorida.ShealsoisaResearchHealthScientistattheBrainRehabilitationResearchCenter,MalcomRandallVA,Gainesville,Florida.