Citation
Noise Robust Algorithms to Improve Cell Phone Speech Intelligibility for the Hearing Impaired

Material Information

Title:
Noise Robust Algorithms to Improve Cell Phone Speech Intelligibility for the Hearing Impaired
Creator:
Ramani, Meena
Publisher:
University of Florida
Publication Date:
Language:
English

Thesis/Dissertation Information

Degree:
Doctorate ( Ph.D.)
Degree Grantor:
University of Florida
Degree Disciplines:
Electrical and Computer Engineering
Committee Chair:
Harris, John G.
Committee Members:
Slatton, Kenneth C.
Van Oostrom, Johannes H.
Holmes, Alice E.
Graduation Date:
5/1/2008

Subjects

Subjects / Keywords:
Auditory perception ( jstor )
Cell phones ( jstor )
Ears ( jstor )
Hearing aids ( jstor )
Hearing loss ( jstor )
Hearing tests ( jstor )
Listening ( jstor )
Loudness ( jstor )
Mutual intelligibility ( jstor )
Signals ( jstor )
acclimitization, aid, cell, fitting, hearing, intelligibility, loss, noise, phone, robust, speech
Genre:
Unknown ( sobekcm )

Notes

Abstract:
Cell phone speech can lead to a difficult listening environment because of the environmental noise, the reduced bandwidth, the packet drop offs and the vocoder artifacts. This is especially true for hearing-impaired listeners who require a 9 dB improvement in signal to noise ratio (SNR) compared to normal-hearing listeners in order to understand speech in noise. This research explored various means to improve cell phone speech intelligibility for the hearing-impaired and resulted in the development of three novel hearing enhancement algorithms. The first algorithm developed by us is the recruitment based compensation (RBC) fitting method. RBC is a hearing enhancement algorithm aimed at improving speech intelligibility (SI) for unaided listeners with sensorineural hearing loss. It is a fitting algorithm which adjusts the gain parameters of the cell phone based on the individual?s threshold of hearing. It provides multiple band gain and compression to make cell phone speech audible and within the reduced dynamic range of the hearing-impaired individual. Subjective hearing in noise tests (HINT) run on hearing-impaired subjects reveal that RBC shows a 15 dB improvement in SNR when compared to linear amplification which is typical of the cell phone volume control. The second algorithm developed by us is the noise robust recruitment based compensation (NR-RBC) algorithm. NR-RBC is derived from RBC but uses the masked thresholds in noise instead of the thresholds in quiet. NR-RBC provides hearing loss compensation and automatic volume control in noisy environments. The objective speech intelligibility index (SII) scores indicate that NR-RBC has high speech usage when compared to all the other fitting methods. Both RBC and NR-RBC received a speech quality mean opinion score (MOS) of 'Good.' Though RBC and NR-RBC were designed with the hearing-impaired in mind the algorithm proves beneficial to the normal-hearing person with slight modifications. This resulted in a 5 dB improvement in SNR when compared to linear amplification and a speech quality rating of 'Good.' For the aided hearing-impaired population, the hearing aid fitting acclimatization method was developed to improve speech intelligibility. Acclimatization occurs because of the plasticity of the auditory cortex. Acclimatization modeling was carried out using neural networks which were trained with multi-session Phonak hearing aid fitting data. This method is to be used in conjunction with existing hearing loss fitting algorithms and predicts the effect of hearing aid acclimatization. The mean square error (MSE) between the predicted values and the optimal values averaged across the parameters is lower than with the initial settings.
Statement of Responsibility:
by Meena Ramani

Record Information

Source Institution:
University of Florida
Holding Location:
University of Florida
Rights Management:
Copyright Meena Ramani. Permission granted to the University of Florida to digitize, archive and distribute this item for non-profit research and educational purposes. Any reuse of this item in excess of fair use or other copyright exemptions requires permission of the copyright holder.
Embargo Date:
5/31/2010
Resource Identifier:
689998376 ( OCLC )

Downloads

This item has the following downloads:


Full Text

PAGE 1

1

PAGE 2

2

PAGE 3

Idedicatethisdissertationtomyincrediblefamilywhohavebeenaconstantsourceofsupportandinspiration. 3

PAGE 4

IwouldliketothankmyadvisorDr.JohnG.Harrisforhisencouragement,patienceandguidance.HetaughtmetoasktherightquestionsandgettotherootoftheproblemandthatissomethingIwillalwaysbegratefulfor.Ialsothankhimformakingthehybridgroupahomeawayfromhomeforallofus.IwouldliketothankDr.AliceE.Holmesformeetingwithmeeveryweekandhelpingmeunderstandthefascinatingeldofaudiology.IalsothankDr.HolmesforaccesstotheShandsspeechandhearingclinicwhereImetamazingpeoplewhofurtherstrengthenedmyresolvetoworkonhearinglosscompensation.IwouldliketothankDr.HansvanOostromandDr.ClintSlattonforbeingpartofmycommitteeandprovidingmewithhelpfulinsights.IwouldliketothanktheMotorolaiDENgroupforfundingtheresearchinChapters 2 and 3 .Overthecourseofmyinter-disciplinaryresearch,Ihadtheopportunitytoworkwithseveralaudiologystudentswhohavehelpedmelookathearinglossfromanon-engineeringperspective.IthankSharonPowell,RyanBaker,ShariKwonandBrittanySakowiczforthat.Ialsothankthemforhelpingmerunthesubjectiveevaluationtestsandforhelpingmecollectthehearingaidttingdata.IfeelextremelyblessedtobepartofthehybridgroupwhereIgettointeractwithbrilliantpeopleonadaytodaybasis.Apartfrombeingextremelyknowledgeableresearchers,theyarealsosomeofthenicestpeopleIhavemet.IthankKwansun,Xiaoxiang,Jeremy,Ismail,Mark,Harsha,Du,ChristyandthemanyothersformakingmyPhDlifeextraspecial.Finally,Iwouldliketothankmyfamilyfortheirunwaveringfaithinme. 4

PAGE 5

page ACKNOWLEDGMENTS ................................. 4 LISTOFTABLES ..................................... 7 LISTOFFIGURES .................................... 8 ABSTRACT ........................................ 11 CHAPTER 1INTRODUCTION .................................. 13 1.1SensorineuralHearingImpairment ....................... 14 1.1.1CausesofSensorineuralHearingLoss ................. 15 1.1.2PerceptualMeasureofSensorineuralHearingLoss .......... 15 1.1.3CharacteristicsofSensorineuralHearingLoss ............. 16 1.1.4ModelingSensorineuralHearingLoss ................. 18 1.2SpeechIntelligibilityandQuality ....................... 19 1.2.1FactorsInuencingSpeechIntelligibilityandQuality ........ 19 1.2.2SpeechIntelligibilityMeasures ..................... 20 1.2.3SpeechQualityMeasures ........................ 21 1.3CellPhoneSpeechIntelligibility ........................ 21 2HEARINGLOSSCOMPENSATIONALGORITHMS ............... 31 2.1ReviewofExistingHearingLossCompensationAlgorithms ......... 31 2.1.1Threshold-OnlyGainPrescriptionProcedures ............ 32 2.1.2SuprathresholdGainPrescriptionProcedures ............. 33 2.2DevelopmentofRecruitmentBasedCompensation .............. 34 2.3ParameterAnalysisofRBC .......................... 36 2.3.1DynamicConstantsofCompression .................. 36 2.3.2FilterBankAnalysis .......................... 38 2.3.3Real-TimeImplementationIssues ................... 39 2.4PerformanceAnalysisoftheRBCAlgorithm ................. 39 2.4.1PerformanceofAlgorithminTermsofSpeechQuality ........ 40 2.4.2PerformanceofAlgorithmintermsofSpeechIntelligibility ..... 41 2.5Summary .................................... 43 3NOISEROBUSTHEARINGENHANCEMENTALGORITHMS ........ 60 3.1EectsofNoiseonCellPhoneSpeech ..................... 60 3.2DevelopmentofNoiseRobustRecruitmentBasedCompensation ...... 61 3.2.1SingleMicrophoneNoiseEstimation .................. 62 3.2.2CalculatingtheNoiseMaskingThreshold ............... 62 3.2.3DerivationofNoiseRobustRecruitmentBasedCompensation ... 63 5

PAGE 6

.............. 64 3.3.1PerformanceofAlgorithminTermsofSpeechQuality ........ 64 3.3.2PerformanceofAlgorithmintermsofSpeechIntelligibility ..... 65 3.4Summary .................................... 66 4ACCLIMATIZATIONMODELINGFORTHEAIDEDHEARINGIMPAIRED 76 4.1DevelopmentoftheFittingSatisfactionScale ................ 77 4.2HearingAidFittingData ........................... 78 4.2.1HearingAidFittingDataCollection .................. 78 4.2.2Multi-SessionHearingAidFittingDataAnalysis ........... 78 4.3ModelingtheAcclimatizationEect ...................... 79 4.4PerformanceAnalysisofModel ........................ 80 4.5Summary .................................... 81 5CONCLUSIONS ................................... 99 APPENDIX ASURVEYOFHEARING-IMPAIREDCELLPHONEUSERS .......... 102 A.1Participants ................................... 102 A.2Results ...................................... 102 A.2.1CellPhoneUsage ............................ 102 A.2.2ElectromagneticInterference ...................... 103 A.2.3CellPhoneSpeechandRingerLevel .................. 103 A.2.4SummaryandConclusions ....................... 103 BCELLPHONEHEARINGEVALUATIONQUESTIONNAIRE ......... 106 CANALYSISOFTHEFOCUSGROUPDISCUSSIONS .............. 110 C.1Participants ................................... 110 C.2FocusGroupMainThemes ........................... 110 C.2.1AidedCellPhoneListeningProblems ................. 110 C.2.2IdealHearingAidCompatibleCellPhone ............... 111 C.2.3CommentsonaCellPhoneAssistiveListeningDevice ........ 111 DPHYSIOLOGYOFHEARING ........................... 112 REFERENCES ....................................... 117 BIOGRAPHICALSKETCH ................................ 123 6

PAGE 7

Table page 1-1Meanopinionscore5pointscale .......................... 23 2-1ThekfconstantforPOGO ............................. 43 2-2ThekfconstantforNAL ............................... 43 3-1Sourcesofcellphonenoiseandnoise-reductionmethods ............. 67 3-2CriticalbandsandFFTbins ............................. 67 4-1Speechintelligibilitybasedttingsatisfactionscale ................ 81 4-2Phonakhearingaidttingparameters ....................... 82 7

PAGE 8

Figure page 1-1Eectsofagingonhearingthresholds ........................ 23 1-2Matlabaudiogramgraphicuserinterface(GUI) .................. 24 1-3Decreasedaudibilitycharacteristicofsensorineuralhearingloss(SNHL) ..... 24 1-4DecreaseddynamicrangecharacteristicofSNHL ................. 25 1-5DecreasedfrequencyresolutioncharacteristicofSNHL .............. 25 1-6DecreasedtemporalresolutioncharacteristicofSNHL ............... 26 1-7Spectrogramsofcellphonespeechfornormal-hearingandsimulatedSNHL ... 27 1-8SimulatedSNHLmodel ............................... 28 1-9Speechintelligibility(SI)asafunctionofbandwidth ............... 28 1-10Hearinginnoisetest(HINT)MatlabGUI ..................... 29 1-11SpeakerresponsefortheMotorolai265 ....................... 29 1-12Meanopinionscore(MOS)speechqualityratingsforcellphonevocoders .... 30 1-13Natureofcellphonehearingproblems ....................... 30 2-1Classicationofexistinghearingaidttingmethods ................ 43 2-2GainsprescribedbytheFig6method ........................ 44 2-3Input-Outputcurveat2kHzobtainedfromthevisualinputoutputlocator ... 44 2-4Variationofdesiredsensationlevel(DSL)prescribedgainwithhearingloss ... 45 2-5Recruitmentbasedcompensationsystem ...................... 45 2-6Computationofgainbasedonloudnessrecruitment ................ 46 2-7Estimateddependenceofrecruitmentrangeonhearingloss ........... 46 2-8Compressioninput-outputandgaincurves .................... 47 2-9Eectofvariationoflterbanksizeonspeechintelligibility ........... 48 2-10AverageMOSscoresforthehearing-impaired ................... 49 2-11AudiogramonthephoneJavamidlet ........................ 50 2-12Audiogramsofallthehearing-impairedlisteners .................. 50 8

PAGE 9

........................... 51 2-14ThePESQobjectivespeechqualityscorefornormal-hearingandhearing-impaired 51 2-15SpectrogramofSNHLandlinear-ampliedspeech ................. 52 2-16Spectrogramofnormal-hearingandlinear-ampliedspeech ........... 53 2-17AverageMOSscoresforthehearing-impaired ................... 54 2-18AverageMOSscoresforthenormal-hearing .................... 55 2-19Speechintelligibilityindex(SII)scoresfornormal-hearingasafunctionofSNR 56 2-20AverageHINTscoresofthehearing-impairedforwidebandspeech ....... 57 2-21AverageHINTscoresofthehearing-impairedforcellphonespeech ....... 58 2-22AverageHINTscoresofthenormal-hearingforcellphonespeech ........ 59 3-1Noiserobustrecruitmentbasedcompensation(NR-RBC)system ......... 68 3-2ThePESQobjectivespeechqualityscoreforvariousHAttingalgorithms ... 69 3-3SpectrogramofSNHLandlinear-ampliedspeech ................ 70 3-4Spectrogramofnormal-hearingandlinear-ampliedspeech ........... 71 3-5AverageNR-RBCMOSscoresforthehearing-impairedlistener ......... 72 3-6AverageNR-RBCMOSscoresforthenormal-hearinglistener .......... 73 3-7TheSIIscoresforsimulatednormal-hearingasafunctionofSNR ........ 74 3-8TheHINTscoresforNR-RBCforhearing-impaired ................ 74 3-9TheHINTscoresforNR-RBCfornormal-hearing ................. 75 4-1ComparisonofClarogainparametersfrominitialtonaltting ......... 83 4-2ComparisonofSaviagainparametersfrominitialtonaltting ......... 84 4-3ComparisonofSaviacompressionparametersfrominitialtonaltting ..... 85 4-4ComparisonofSaviacompressionparametersfrominitialtonaltting ..... 86 4-5PhonakSaviamaximumtrendinttingparametervariation ........... 87 4-6PhonakClaromaximumtrendinttingparametervariation ........... 87 4-7PhonakExtramaximumtrendinttingparametervariation ........... 88 4-8PhonakValeomaximumtrendinttingparametervariation ........... 89 9

PAGE 10

........... 90 4-10PhonakPerseomaximumtrendinttingparametervariation .......... 91 4-11StructureoftheMLPusedtomodelmulti-sessionttingtrends ......... 92 4-12PhonakSavianeuralnetworkmodelingresultsfor40dBgainparameter ..... 93 4-13PhonakSavianeuralnetworkmodelingresultsfor60dBgainparameter ..... 94 4-14PhonakSavianeuralnetworkmodelingresultsfor80dBgainparameter ..... 95 4-15PhonakSavianeuralnetworkmodelingresultsforCRparameter ......... 96 4-16PhonakSavianeuralnetworkmodelingresultsforTKparameter ........ 97 4-17PhonakSavianeuralnetworkmodelingresultsforMPOparameter ....... 98 A-1Degreeofhearingimpairment ............................ 104 A-2Degreeofhearingimpairmentforsurveyparticipants ............... 105 D-1Structureofthehumanear ............................. 114 D-2Organofcorti ..................................... 115 D-3Electronmicrographoftheorganofcorti ...................... 115 D-4Frequencysensitivityofthebasilarmembrane ................... 116 10

PAGE 11

Cellphonespeechcanleadtoadicultlisteningenvironmentbecauseoftheenvironmentalnoise,thereducedbandwidth,thepacketdroposandthevocoderartifacts.Thisisespeciallytrueforhearing-impairedlistenerswhorequirea9dBimprovementinsignaltonoiseratio(SNR)comparedtonormal-hearinglistenersinordertounderstandspeechinnoise.Thisresearchexploredvariousmeanstoimprovecellphonespeechintelligibilityforthehearing-impairedandresultedinthedevelopmentofthreenovelhearingenhancementalgorithms. Therstalgorithmdevelopedbyusistherecruitmentbasedcompensation(RBC)ttingmethod.RBCisahearingenhancementalgorithmaimedatimprovingspeechintelligibility(SI)forunaidedlistenerswithsensorineuralhearingloss.Itisattingalgorithmwhichadjuststhegainparametersofthecellphonebasedontheindividualsthresholdofhearing.Itprovidesmultiplebandgainandcompressiontomakecellphonespeechaudibleandwithinthereduceddynamicrangeofthehearing-impairedindividual.Subjectivehearinginnoisetests(HINT)runonhearing-impairedsubjectsrevealthatRBCshowsa15dBimprovementinSNRwhencomparedtolinearamplicationwhichistypicalofthecellphonevolumecontrol.RBCalsoshowsa6dBimprovementinSNRwhencomparedtothedesiredsensationlevel(DSL)ttingmethodwhichisapopularaudiologyoption. 11

PAGE 12

ThoughRBCandNR-RBCweredesignedwiththehearing-impairedinmindthealgorithmprovesbenecialtothenormal-hearingpersonwithslightmodications.Thisresultedina3dBimprovementinSNRwhencomparedtoDSLusingRBC,a13dBimprovementinSNRusingNR-RBCandaspeechqualityratingof\Good." Fortheaidedhearing-impairedpopulation,thehearingaidttingacclimatizationmethodwasdevelopedtoimprovespeechintelligibility.Acclimatizationoccursbecauseoftheplasticityoftheauditorycortex.Acclimatizationmodelingwascarriedoutusingneuralnetworkswhichweretrainedwithmulti-sessionPhonakhearingaidttingdata.Thismethodistobeusedinconjunctionwithexistinghearinglossttingalgorithmsandpredictstheeectofhearingaidacclimatization.Themeansquareerror(MSE)betweenthepredictedvaluesandtheoptimalvaluesaveragedacrosstheparametersislowerthanwiththeinitialsettings. 12

PAGE 13

Thesenseofhearingplaysapivotalroleinhumaninteractionandcommunication.Acousticpressurewavesaretransducedbythecochleaintoelectricalneuralsignalswhichareprocessedbythebraintoprovideameaningfulcognitiveexperience.Hearingimpairmentcanreducetheabilitytocommunicatesuccessfully.Theinabilityofbeingabletounderstandwhatisbeingsaidcanresultinsocialandemotionalisolation[ 1 ].Thetelephone,oneofthemostimportantinventionsofthe19thcentury,wastheresultofAlexanderGrahamBell'sworkoncommunicationdevicesforthehearing-impaired.Telephoneshavenowbecomeanintegralpartofhumancommunicationandprovideeasymeansoflong-distancecommunication.Theinventionofthewirelesscellphonehasfurtherleadtoaneaseincommunication.CellphonesarethemoderndaySwissArmyknivesandarepackedwithamyriadofhardwareandsoftwarefunctionalities.AsofJune2007,thereare243million[ 2 ]cellphonesubscribersintheUnitedStatesandthisnumberisgrowing. IntheUnitedStatesalonethereare28million[ 3 ]peoplewhoarehearing-impaired.Yetlessthan8%ofthemusehearingaidsthoughtheycouldobtainsignicantimprovementwiththem.Thisismainlybecauseofthehighcostsandthestigmaattachedtousinghearingaids.Studieshaveshownthathearing-impairedlistenersrequirea9dBimprovementinsignaltonoiseratio(SNR)[ 4 ]whencomparedtonormal-hearinglistenersinordertounderstandconversationalspeech.Hearingaidscanhelpsatisfythisrequirementtoacertainextent.Unfortunatelyhearingaidsandcellphonesarenotcompletelycompatiblebecauseofelectromagneticinterference(EM)[ 5 ].Theamountofinterferencedependsontheamountofradio-frequency(RF)emissionproducedbytheparticularcellphoneandtheimmunityoftheparticularhearingaid.TheIEEEC63.19standard[ 6 ]providesaratingscalewhichservesasameasureofthecompatibilitybetweencellphonesandhearingaids.Consumerscanlookforthisratingwhilepurchasingacell 13

PAGE 14

A hastheresultsofasurveyconductedatUniversityofFloridawhichindicatesthatinordertoavoidtheEMinterference,mostaidedhearing-impairedlistenersprefertoremovetheirhearingaidinordertousethecellphone. Cellphonespeechcansometimesbediculttounderstand,becauseoftheenvironmentalnoise,thereducedsignalbandwidth(300{3400Hz),thepacketdroposandthevocoderartifacts.Theenvironmentalnoisemasksthespeechwhilethereducedsignalbandwidthandvocoderartifactsresultinalossinnaturalnessandintelligibility[ 7 ].Hearing-Impairedlistenersoftenndcellphonesspeechtobeunintelligible.Modernhearingaidsareextremelylowpowerdigitalsignalprocessor(DSP)basedsystemsandprovidegainandcompressionbasedontheindividualshearinglossthroughaprocessreferredtoashearing-aidtting.Inaddition,hearingaidDSPsalsorunfeedbackcancellationandnoisereductionalgorithms.Inordertoimprovecellphonespeechintelligibilityforthehearing-impaired,powerfulhearingenhancementalgorithmscanberunonthecellphones.Thischapterwillprovideanintroductiontosensorineuralhearing-impairmentandcellphonespeechintelligibility. 8 ].Becauseofthenonlinearnatureofthisloss,simplelinearamplicationwillnotrestorenormalhearing.Hearinglosscanbecategorizedbasedontheseverityasmild(25{40dBHL),moderate(40{70dBHL),severe(70{95dBHL)andprofound(95dBHL.)Hearingaidscanhelppeoplewithmildtoseverehearingloss.ThehearingaidalgorithmsattempttoimitatetheOHCsactingtoreplacethedamagedor 14

PAGE 15

9 ].CochlearimplantshavetobeusedifthereissignicantIHClossasisthecasewithprofoundhearingloss.Appendix D providesashortdescriptionofhowwehearanddescribestherolesoftheIHCsandtheOHCs. 1-1 showstheeectsofagingonthethresholdsofhearing.Presbycusisoccursduetowearandtearofthehaircellsofthecochlea.Lossesupto60dBHLcanbeassumedtobecausedbydamagetotheOHCs.Forlossesgreaterthan80dBHL,boththeIHCsandOHCshavetobedamaged. Sensorineuralhearinglosscausedduetoexposuretoloudsoundsiscallednoiseinducedhearingloss(NIHL)[ 10 ].Soundsathighintensitiesfatiguethehaircellsofthecochleaanddependingonthedurationofexposurethismaycausepermanentdamage.PortableaudiodeviceslikeiPodscanproducesoundlevelswhichcancauseirreversibledamageevenwhenplayedforacoupleofminutes[ 11 ].Cellphonesandbluetoothheadsetsalsoproducesoundlevelswhichcancauseconsiderabledamage.Recentlytherehasbeenalotofeortonthepartofportableaudiodevicemanufacturerstoeducatethepubliconsafelisteningpractices.Safelisteninglevelsformusicandspeechhavebeenestimated[ 11 ]usingexistingnoiseexposurestandards[ 12 ],[ 13 ]. 14 ].Theaudiogrammeasuresthethresholdofhearinginquiet.Itisobtainedbyplayingpuretonesornarrowbandsofnoise,typicallybetween250{8000Hz,atvariousintensitylevelstillitisjustaudible. ThethresholdsofhearingthusobtainedarecomparedtotheaveragenormalhearingthresholdsandthedierenceisreportedindBHL(hearinglevel).Peoplewith`perfect'hearingwillhaveanaudiogramof0dBHL.Normalhearingisdenedashavingallpoints 15

PAGE 16

1-2 showsanaudiogramofapersonwithamildhearinglossmeasuredusingaMatlabGUI.Onanaverage,apuretoneaudiogramtakes5minutestobemeasured. 15 ],[ 8 ]. D describesthemechanicsbehindhowwehearandhowhearinglossoccurs.Figure 1-3 showsthehearingthresholdsforahearing-impairedandanormal-hearinglisteneranditcanbeseenthatthehearing-impairedlistenerhashigherthresholdsofhearingespeciallyathighfrequencies. Thisdecreasedaudibilityresultsinlowspeechintelligibilitybecausetheconsonantsandthesecond,thirdformantsofspeechwillnotbeaudible.Sincetheloudnessofspeechisdominatedbythelowfrequencycomponents,thehearing-impairedlistenersdonotrealizethattheyaremissingoutonpartofthesignal[ 16 ].Eventhoughcellphonespeechisbandlimitedto300{3400Hz,for90%ofhearing-impairedlistenersthedegreeofhearinglossworsensfrom500Hz{4kHz[ 17 ]andthisdetrimentallyaectsthecellphonespeechintelligibility.Theaudiogramprovidesadirectmeasureofthedecreasedaudibilityandisusedinallhearingaidttingalgorithms.Decreasedaudibilitycanbecompensatedbyprovidingafrequencydependentgain. 16

PAGE 17

18 ].Thisresultsinadecreaseddynamicrangeofspeechandthisphenomenoniscalledloudnessrecruitment.Loudnessrecruitmentismeasuredusingloudnessscalingexperiments.Figure 1-4 showstypicalloudnessgrowthcurvesmeasuredusingasixpointloudnessscaleforanormal-hearingandahearing-impairedlistener.Decreaseddynamicrangecanbecompensatedbyprovidingcompression. 19 ]referstothedecreaseinfrequencysensitivityandfrequencyselectivity[ 20 ].TheOHCsincreasethesensitivityofthecochleatotheparticularfrequencythattheportionofthebasilarmembraneistunedto.WhentheOHCsaredamagedthissensitivitydecreases.Frequencyresolutioncanbemeasuredusingpsychoacoustictuningcurves.Psychoacousticcurvesaremeasuredbyplayinganaudiblepuretone(probe)andvaryingthelevelofanarrowbandofnoise(masker)tillthetoneisbarelyaudible.Figure 1-5 showsthepsychoacousticcurvesforanormal-hearingandhearing-impairedlistenerfora4kHztonewitha40dBmasker. Thetuningcurveforthehearing-impairedlistenerisatandbroad(Figure 1-5 )[ 21 ].Becauseofthis,thehighenergy,lowfrequencypartsofspeechwillmaskmoreoftheweakerhighfrequencycomponents.Thisisknownasupwardspreadofmasking[ 22 ].Mostoftenenvironmentalnoiseislowfrequencyandbecauseofupwardspreadofmaskinghearing-impairedlistenershaveadiculttimeunderstandingspeechinnoise.Also,ithasbeenshownthatathighintensitylevelsevennormal-hearinglistenershavepoorfrequencyresolution.Thisisbecauseofsaturationofthehaircells.Hearing-Impairedlistenersalwayslistentohighintensitysounds.Thisfurtherworsenstheirfrequencyresolution[ 23 ].Decreasedfrequencyresolutioncanbecompensatedfortoacertainextentbyusingsharpandnarrowlterbankswhileprocessingthespeech. 17

PAGE 18

24 ].Whilelisteningtospeechinannoisyenvironment,normal-hearinglistenersextractmostinformationfromspeechwhenthenoiseislowinmagnitude.Butbecauseofreducedtemporalresolution,thesespeechregionswillbemaskedforthehearingimpaired[ 25 ].Temporalresolutionismeasuredusingpsychoacoustictuningcurves.Figure 1-6 showsthepsychoacousticcurvesforthenormal-hearingandhearing-impairedfora4kHzprobetonewitha40dBmasker.Decreasedtemporalresolutioncanbecompensatedbyvaryingthegainsoastogetnormalmaskingthreshold. 26 ],[ 27 ],[ 28 ]helpinthedevelopmentandtestingofcompensatorytechniques.Forourresearch,weusedamodelbasedonbothMoore[ 26 ]andDuchnowski[ 28 ].Themodelsimulatesthedecreasedaudibilityandtheloudnessrecruitmentaspectsofhearingloss.Spectrogramsofcellphonespeechatanormalconversationallevelforbothnormal-hearingandatypicalmildtosevereSNHLhearinglossof[102030608090]dBHLareshowninFigure 1-7 Thehighfrequencyconsonantinformationofspeechiscompletelymissingforthehearing-impairedandthisresultsinlowspeechintelligibility(Figure 1-7b ).Thehighenergylowfrequencypartofspeechisstillpresentandmakesthespeechaudiblebutunintelligible.Figure 1-8 isthesetupusedtomodelthehearingloss. ThealgorithmusesmultiplelterbandsandcalculatestheHilberttransformforeachlterbandoutput.TheenvelopeofthebandlimitedspeechobtainedfromtheHilberttransformisthenraisedandsmoothedtoobtaintheeectofloudnessrecruitment.Themodiedenvelopeisthenmultipliedwiththenestructurewithintheoriginalenvelope,togeneratesimulatedlossyspeechforthatband.Theoutputsofallthelterbandsarenallysummedtogethertogetthesimulatedlossyspeech.TheMatlabsimulationused30lterbankswithcenterfrequenciesequallyspacedinmelfrequencybetween100Hzto8000Hz. 18

PAGE 19

29 ]andspeechqualityindicateswhetherthespeechmeetstheexpectationsofthelistener.Subjectivemeasuresofevaluatingspeechintelligibilityandqualityarebasedonscoresobtainedvialisteningexperiments.Objectivemeasuresofintelligibilityandqualityrelyonsignal-to-noisemeasurementsandmodelsofhumanspeechperception. Bandwidth 30 ].Measurementsshowthattheintelligibilityofspeechdecreaseswithdecreasingbandwidth.Itisalsoimportantforthefrequencyresponsetobereasonablyatthroughoutit'srange.Forsinglewordsnarrowband(NB)speechyieldsanaccuracyofonly75%,whilewideband(WB)speechresultsina97%accuracy[ 31 ].Thislossofintelligibilityincreaseswhenmultiple-wordspeechsoundsareusedtotestintelligibility(Figure 1-9 ). 32 ].Onlynoisewhichfallswithinthesamecriticalbandwidthasspeechcancontributetothemaskingofspeech.Environmentalnoiseispredominantlylowfrequencyandisastrongmaskerwhichathighsoundpressurelevelscanmaskboththespeechvowelsandconsonants[ 33 ],[ 34 ]. 35 ].Independentmulti-bandoperationschangethetemporalandspectralenvelopeofspeechandthisdetrimentallyaectsthespeechcuesresultinginlowSI[ 36 ].Toavoidaudibleartifacts,multi-bandtechniquesareusuallyfollowedbysomepost-processinglikeenvelopesmoothening. 19

PAGE 20

37 ].Listenersareplacedina65dBAconstantnoiseenvironmentandspeechsentencesatvarioussignallevelsarepresentedtothemviaheadphones.Thelistenerthenhastorepeatwhatheheard.Theintensityofthenextsentenceisadaptivelyvariedby2dBor4dBbasedontheirresponse.Itisstipulatedthatafter10sentences,thenalsentenceintensitylevelconvergestoalevelatwhichthelistenerrecognizes50%ofthesentencescorrectly.Thismethodofscoringintelligibilityiscalledthereceptionthresholdforsentences(RTS).AMatlabGUIwasusedtoautomatethetest(Figure 1-10 ).TheresultoftheHINTisaSNRvaluebasedonRTS.ThelowertheSNRvalue,thehigherthespeechintelligibility. 1{1 Inthisformula,nreferstothenumberoffrequencybandsusedwhichcanvaryfrom6octavebandsto21criticalbands.IiisthebandimportancefunctionandAireferstothebandaudibility,whichrangesfrom0to1andindicatestheproportionofspeechcuesthat 20

PAGE 21

38 ]. 1-1 .AMatlabGUIwascreatedtoautomatetheMOStest. 39 ]recommendedstandardfortheobjectivemeasurementofthespeechqualityofnarrowbandsystems.PESQcomparestheoriginalsignaltoamodiedversionofthesameandpredictstheperceivedqualitythatwouldbegiventothemodiedsignalbysubjectsinasubjectivelisteningtest.TherangeofthePESQscoreis-0.5(extremelylowquality)to4.5(excellentquality). 7 ].Figure 1-11 showsthefrequencyresponseoftheMotorolai265cellphoneloudspeaker.Itcanbenotedthatresponseisnotatacrossfrequenciesandthisfurtherresultsinalossinintelligibility. 21

PAGE 22

Inaddition,cellphonespeechalsohasvocoderartifacts.Basicallyforanyvocoder,theinputspeechisrstdividedintooverlappingframes.Asetofmodelparametersarethenestimatedforeachframe,quantizedandthentransmitted.Atthereceiver,thedecoderreconstructsthemodelparametersandusesthemtogenerateasyntheticspeechsignal.Theadvancedmulti-bandexcitation(AMBE)vocoderiscommonlyusedwithMotorolahandsets[ 40 ].Figure 1-12 showstheMOSspeechqualityratingsforthemostcommonlyusedvocoders.Dependingonthedatarate,AMBEhasanaverageMOSscoreof3.2to3.7. ClarityandtheEARfoundationconductedaresearchstudyamongarandomgroupof458babyboomersbetweentheageof41-60[ 41 ].53%ofthebabyboomersreportedhavingatleasta`mild'lossandover57%ofbabyboomershadtroublehearingontheircellphones.40%ofthosewhohadproblemsusingthecellphonesaidtheywouldusethecellphonemoreofteniftheycouldheartheconversationsmoreclearlywhileusingit.Figure 1-13 liststhenatureofthecellphonehearingproblems. Inordertobetterunderstandthecellphoneneedsofthehearing-impaired,focusgroupsandsurveysoncellphonehearingwerecarriedoutattheUniversityofFlorida.A`Cellphonehearingevaluation'questionnaire,availablefromAppendix B ,wascreatedandhandedoutto84patientsattheShandsspeechandhearingclinicinGainesville.Appendix A hastheresultsfromthequestionnairebasedsurveyandAppendix C discussesthemainthemesobservedatthetwofocusgroupswhichwereconducted.Theresultsfromboththeseindicatethenecessityofhavingalgorithmsrunonthecellphoneinordertoenhancethehearingandimprovespeechintelligibility. 22

PAGE 23

Meanopinionscore5pointscale MOSQuality Figure1-1. Eectsofagingonhearingthresholds[ 42 ] 23

PAGE 24

Figure1-3. Thresholdsofhearingforthenormal-hearingandhearing-impaired 24

PAGE 25

Loudnessgrowthcurveforthenormal-hearingandhearing-impaired Figure1-5. Psychoacoustictuningcurveshowingdecreasedfrequencyresolution 25

PAGE 26

Psychoacoustictuningcurvesforthenormal-hearingandhearing-impaired 26

PAGE 27

b SpectrogramsofcellphonespeechforA)Normal-HearingB)MildtoSevereSNHL 27

PAGE 28

Simulatedsensorineuralhearinglossmodel Figure1-9. Speechintelligibilitymeasuredusingarticulationindexasafunctionofbandwidth[ 31 ] 28

PAGE 29

HearinginNoiseTestMatlabGUI Figure1-11. SpeakerresponsefortheMotorolai265 29

PAGE 30

MOSspeechqualityratingsforcellphonevocoders[ 40 ] Figure1-13. Natureofcellphonehearingproblems[ 41 ] 30

PAGE 31

Hearingaidshavetobecustomizedtoeachuser'suniquehearingloss.Thisisachievedbyadjustingthegainandcompressionvaluesofthehearingaiddigitalsignalprocessor(DSP)usingaprescriptivealgorithm.Thisprocessisreferredtoashearingaidttingandtheprescriptivealgorithmusediscalledthehearinglosscompensationalgorithmorthehearingaidttingalgorithm.AsmentionedinChapter 1 andinAppendix D ,mildtomoderatelyseveresensorineuralhearingloss(SNHL)isprimarilycausedbydamagetotheouterhaircellsofthecochlea.Soineect,thehearinglosscompensationalgorithmhastoimitatetheouterhaircells(OHC)[ 9 ].Inordertorunhearingenhancementalgorithmsonthecellphoneforthehearing-impaired,theDSPofthephonehastobettothelistener'shearingloss.Thischapterwillprovideabriefreviewoftheexistingttingalgorithmsandwilldetailthedevelopmentofanewhearinglosscompensationalgorithmforcellphonespeech,therecruitmentbasedcompensation(RBC)method.Speechprocessedbythenewalgorithmwillbeshowntohavehigherintelligibilityandqualitythantheexistingmethods. 15 ]whichvaryintheirrationalebehindgainprescription.Somealgorithmsprescribegainsothatthespeechisalwaysatamostcomfortablelevel(MCL)[ 43 ],othersuseloudnessnormalizationorloudnessequalization[ 44 ]astherationale.Loudnessnormalizationisameansofprescribinggainsoastomaketheloudnessgrowthcurveofthehearing-impairedthesameasthatfornormal-hearing.Loudnessequalizationisbasedontheprincipleofequalizingtheloudnessinformationacrossfrequencies.Intelligibilityisassumedtobemaximizedwhenallthebandsofspeechareperceivedtohavethesameloudness[ 45 ].Figure 2-1 showsthebasicclassicationofthehearingaidgainttingalgorithms.AllthesealgorithmshavebeenimplementedinMatlab. 31

PAGE 32

46 ].Justmirroringtheaudiogramwouldresultinanineectivettingsincetheoutputwillreachuncomfortableloudlevelswhentheinputsignalishighinintensity.ThisisbecauseofthedecreaseddynamicrangeaspectofSNHL.Sincethreshold-onlyalgorithmsdonotincludecompressionintheprescription,theyshouldbefollowedbyoutputlimitingcompressiontopreventthesoundsfromgettingtooloud. 2{1 IGf=0:5Hf(2{1) HereIGfisthegainandHfisthefrequencydependenthearingloss, 47 ]isa1 2gainrulewithanattenuationtermatthelowfrequencies.Thisisdonetodecreasetheupwardspreadofmasking.TheformulaforttingisgivenbyEquation 2{2 IGf=0:5Hf+kf(2{2) HereIGfistheGain,Hfisthefrequencydependenthearingloss,andkfisasshowninTable 2-1 .POGOcanbeusedforhearinglossesupto80dBHL. 48 ]ofAustraliapublishedthenationalacousticlab-revised(NAL-R)formulain1983.Itisthemostpopularofthethreshold-onlybasedttingmethods.TheaimoftheNAL-RprocedureistomaximizelistenerintelligibilityattheMCLbyequalizingloudness.TheNAL-RttingformulaisgivenbyEquation 2{3 .Table 2-2 indicateshowtheconstantkfvarieswith 32

PAGE 33

49 ]procedurefollowstheloudnessnormalizationrationaleformediumandhighlevelinputsignals.Fig6prescribesgainsforthreedierentinputintensitylevels(40dBSPL,65dBSPLand95dBSPL)basedontheaudiogramandaverageloudnessgrowthdata.Thethreelevelsofspeechrepresentthedierentlevelsofconversationalspeechwith40dBSPLrepresentingsoftspeech,65dBSPLrepresentingconversationalspeechand95dBSPLloudspeech. Figure 2-2 showsthetargetsasprescribedbyg6.The95dBSPLcurveprovideslittlegainforthelowfrequencysoundswhicharemoreintensethanthehighfrequencysoundsevenforconversationallevelspeech.The65dBSPLand40dBSPLcurvesprovidesmoregainatthehighfrequencies.Itcanbeseenthattheamountofgaindecreaseswhentheinputlevelincreases. 50 ]techniqueisbasedonloudnessnormalizationandusesloudnessscalingexperimentsinsteadofaverageloudnessgrowthcurves.Theloudnessscalingprocedureusedisthecontourtestandinvolvesplayingpulsedwarbletonesinascendingorderfrom5dBtillthesubjectindicatesthatthestimulusisattheMCL. Ateachlevelthesubjectusesa7pointratingscaletodescribeitsloudness.Thesevenloudnesscategoriesforwarbletonesarecondensedtothreecategoriesforspeech 33

PAGE 34

2-3 .Thevisualinputoutputlocator(VIOLA)programthenplotsforeachfrequencyaninput-outputcurvewith2compressionthresholdsand2compressionratios.Anexampleinput-outputgraphisshowninFigure 2-3 .Thediagonallineacrossthegraphrepresentsthe0dBgain.ThedistancebetweentheIHAFFprescribedtargets(theasterisks)andthediagonallineisthegaintoamplifysoft,averageandloudinputspeechforthatfrequency. 51 ],[ 52 ]aimsatmakingspeechcomfortablyloudandaudible.ThegainfordierenthearinglossandfrequencyasusedintheDSL4.0computerprogramisshowninFigure 2-4 .ThecompressionratioprescribedbyDSLislargerthanthatrequiredtonormalizeloudnessandisprescribedsoastottheextendeddynamicrangefromthenormal-hearingthresholdtothehearing-impairedUCLintothereduceddynamicrangeofthehearingimpaired.DSListhemostpopularsuprathresholdttingmethod. 53 ].ForournovelmethodthisisachievedbyusinglterbanksasshowninFigure 2-5 .HereS(n)istheincomingcellphonespeechsignalwhichistobeenhanced.Processingiscarriedoutinthetimedomainusingaframe-by-frameapproach.S(n)isfedtoalterbankwhichhas14bandswithcenterfrequenciesequallyspacedinmelfrequencybetween300-3400Hz.Thegaincomputationblockusestheenergyperbandandtheuser'shearingthresholdstoprescribeagainandcompressiontermforeachbandaspertheRBCformula.ThesignalsfromeachbandarenallycombinedtogethertogettheenhancedspeechsignalSe(c).TheRBCgainblockshouldbefollowedbyaoutputlimitingblockwhichmakessurethatthesoundsnevergetpainfullyloud.Thecompressionratioandthresholdforthisstagearexedat:10:1and110dBSPLrespectively.Loudnessnormalizationisamethodofprescribinggainssoastomaketheloudnessgrowthcurveforthehearing-impairedthesameasthatforthenormal-hearing.Figure 2-6 showsthe 34

PAGE 35

In1959,HallpikeandHood[ 54 ]showedthattherangeofrecruitment,TcTi,isafairlyorderlyfunctionofthehearingloss,TiTnandisindependentoffrequencyforunilateralhearingloss.Miskolczy-Fodor[ 55 ]furtherreportedthisbehaviorforpresbycusis.BoththeserelationshipsareasshowninFigure 2-7 Letbedenedastheanglebetweentherecruitmentcurveandthehorizontalaxis(Figure 2-6 ).TherelationbetweenandhearinglosscanbedescribedbyEquation 2{4 HerehearinglossHL=TiTn.IfR=TcTiisusedtorepresenttherecruitmentrange,therelationbetweentherecruitmentrangeandhearinglosscanbedescribedbyEquation 2{5 Inordertoachieveloudnessnormalization,thealgorithmampliesthesignalineachchannelsuchthattheoutputlevelisrelatedtotheinputlevelbythesolidline.Asthelevelisincreased,thegaindecreasesuntilatTcthegainbecomesone.Fromtheaudiogram,wecancomputeandRandhencethegainfactorperchannelcanbecomputed.Thisapproachwhichisbasedonthefrequencyindependentrelationshipbetweenrecruitmentandhearinglossiscalledtherecruitmentbasedcompensation(RBC)method. 35

PAGE 36

2{6 tan(w)1=47+0:45HL(w) (2{6) HereHL(w)isthehearinglossatthecenterfrequencyofeachbandwhichisobtainedbythelinear-interpolationoftheaudiogram.TheRBCalgorithmincludescompressionaspartoftheprescriptionandthecompressionratioforeachchannelisgivenbyCR(w)=1 Whiletheexistingalgorithmswhicharealsobasedontheconceptofloudnessnormalizationrequiretheloudnessgrowthcurveforeachfrequencyortheaverageloudnessgrowthcurves,alltheRBCmethodrequiresistheaudiogramofthehearing-impairedperson. 2.3.1DynamicConstantsofCompression 56 ].Compressioncanbecarriedouteitherinasinglebandoninmultiplebands.Inmulti-bandprocessing,eachbandusuallyhasdierentcompressioncharacteristicsandthedegreeofcompressioneitherincreasesordecreaseswithfrequency.Typically2or3bandsareused.Increasingthenumberofcompressionbandsbeyond3canresultinaudibledistortion.Fittingalgorithms 36

PAGE 37

Theattackandreleasetimesarethedynamicconstantsofcompressionandspecifyhowquicklyacompressoroperates.Iftheeectofcompressionisinstantaneousaudibleartifactsareproducedbecauseofthesuddenchangeinlevels.ANSIS3.22denestheattacktimeasthetimetakenfortheoutputtostabilizewithin3dBofitsnallevelaftertheinputchangesfrom55to90dBSPL.Thereleasetimeisdenedasthetimetakenfortheoutputtostabilizewithin4dBofitsnallevelaftertheinputfallsfrom90to55dBSPL.Experimentallyanattacktimeof6msandareleasetimeof20mswasfoundtobeideal.TheimplementationoftheattackandreleasetimeconstantscompressionisgivenbyEquations 2{7 (2{7) HereGave(w)istheaveragesmoothedgainperband,Gi(w)istheinstantaneousgainperband,attackandreleasearetheattackandreleaseconstantsasdenedinANSIS3.22. Compressionratioistheinverseoftheslopeoftheinput-outputcurve.Thecompressionratiousuallyvariesfrom1.1:1to3:1.CompressionthresholdistheSPLabovewhichcompressionkicksin.Ifloudnessistobenormalizedcompletely,compressionshouldkickinfromthethresholdofnormal-hearingwhichis0dBSPL[ 57 ].Butusefulspeechsoundsrarelyoccurbelow30dBSPL.Whenthecompressionthresholdis>50dBSPLitistermedashigh-thresholdandwhenthecompressionthresholdis<50dBSPLitistermedaslow-threshold.Widedynamicrangecompression(WDRC)referstosystemswhichhavelowthreshold.Figure 2-8 showsatypicalWDRCcharacteristicswithoutputlimiting. Tillthecompressionthresholdof50dB,thegainislinear.CompressioniseectivefromafterthethresholdtillthethresholdofcompleterecruitmentTc(80dBinthis 37

PAGE 38

2-9 ). InordertouseRBCforthenormal-hearingpopulationalterbanksizeof5wasfoundtobeoptimalasaresultofsubjectivelisteningtests.Thisisbecausenormal-hearinglistenerscanheartheartifactscausedbecauseofmulti-bandgain.Also,sincetypicalnormal-hearingpopulationhavelittleornoloudnessrecruitmenteects,thecompressionparametersvariedfrom1.1to1.5. 58 ]canbeusedclinicallytodetectdeadregionsofthecochlea.Itissimilartomeasuringthresholdsofhearinginnoiseandmeasuresthe 38

PAGE 39

Audiogramonthephone 59 ]wascreatedtomeasurethethresholdsofhearingusingthecellphone.Sincethecellphoneistobeusedasanassistivelisteningdeviceforcellphoneconversationsandnotahearingaid,calibrationisnotakeyissue.Themidletplaystonesatdierentlevelsandthelistenerpressesakeytoindicatehavingheardthesound.TheMotorolaRokerE2has7volumestepsandbyplayingscaledtonewavelesavolumerangefrom3{65dBwasachieved.Figure 2-11 showsadepictionofhowtheaudiogramonthephonewouldlook. 60 ],theDSLmethod,theHGmethod,thePOGOmethodandtheNAL-Rmethod.ThespeechdatabaseunlessotherwisementionedisthestandardHINTdatabase.Cellphonespeechwasobtainedbybandlimitingthespeechto300{3400HzandthenpassingitthroughanAMBEvocoder/decoderblocktointroducedthevocodereects. 2-12 showstheaudiogramsofallthehearingimpairedsubjects.Outputlimitingcompressionwasprovidedforallthealgorithmswithacompressionratioof10:1andacompressionthresholdof110dB. 39

PAGE 40

2-13 wasused.Typicalmildtoseveresensorineuralhearinglossandtypicalnormal-hearingweresimulatedusingtheMatlabhearinglosssimulationblock.Theaudiogramsusedwere:[102030608090]dBHLand[51015202020]dBHLrespectively.Theunprocessedcellphonespeechwaspassedthroughthehearinglossblocktogeneratesimulatedlossspeech.Speechpreprocessedbythevariousttingalgorithmswerepassedthroughthehearinglossblocktogeneratecompensatedspeech. TheobjectivePESQscoreswereobtainedusingtheoriginalcellphonespeechasthereferencesignalandcomparingittoboththesimulatedlossspeechandthecompensatedspeech(Figure 2-14 ). PESQissensitivetodistortionduetocompression.ThetypicalmildtosevereSNHLmodeledherewouldprovideoutputlevelsathighfrequencieswhichwouldturnontheoutputlimitingcompression.ThisresultsinlowPESQscores.IfwecomparewithallthecompressionbasedsystemsRBCdoesthebestfollowedbyDSLandNAL-R.ThescoresalsorevealthatRBCoutdidlinearamplicationandtheotherttingalgorithmsfornormal-hearingsubjectswithanaveragePESQscoregreaterthan\4-Good." 2-15 ).Thesimulatedhearinglossblockwasusedtogeneratethespeech(Figure 2-13 ). Forthehearing-impaired,alotofhighfrequencyinformationismissing(Figure 2-15a ).Linearamplicationdoesnothelpbecauseofthereduceddynamicrangeaspectofthe 40

PAGE 41

2-15b ).Compressionresultsinmorehighfrequenciesandthishelpsimproveintelligibility(Figure 2-15c ). Thespectrogramsforthetypicalnormal-hearingsimulatedspeech,linearlyampliedspeechandthespeechcompensatedusingtheRBCmethodwerealsoobtained(Figure 2-16 ).Thereismorehighfrequencyinformationbecauseoffrequencydependentgainandthishelpsimproveintelligibility. Theseresultsshowthatforboththenormal-hearingandthehearing-impairedRBChasbetterspeechqualityandmoreusefulfrequenciesthanwithjustalineargainwhichiswhatthecellphonesvolumecontroldoes. 2-17 showstheaverageoftheMOSscoresforthehearing-impaired.Forthehearing-impaired,RBChasanaverageMOSscoregreaterthan\4-Good." Figure 2-18 showstheaverageoftheMOSscoresforthenormal-hearing.Forthenormal-hearing,RBChasanaverageMOSscoregreaterthan\4-Good." 3-7 showsthevariationofSIIwithSNRfrom-30to30. 41

PAGE 42

2-20 showstheaveragedSNRforthe10hearing-impairedsubjects,withreferencetothebaseline(lineargain)forwidebandspeech.ThesescoresshowthatRBCdoesthebestfollowedbyDSLandhalf-gain.NAL-Randhalf-Gain.WhencomparedtothelineargaintechniqueRBCprovidesupto15dBimprovementinSNR.ThedierencebetweenRBCandDSLforwidebandspeechisabout3dB. Figure 2-21 showstheaveragedHINTresultswithnarrowbandcellphonespeechinput.ThesescoresindicatethatRBCdoesthebestfollowedbyNAL-Randhalf-gain.Half-Gainprescribesahighergainthanallthettingmethodsbeingtested.Forloudinputlevels,thiswillleadtoadecreaseinintelligibilitybutintheHINTthelevelofspeechisreducedsothegainincrementhelpshalf-gaindobetter.ThedierencebetweenRBCandlineargainis15dB.ThedierencebetweenRBCandDSLforcellphonespeechisabout6dB. Figure 2-22 showstheaveragedHINTresultswithnarrowbandcellphonespeechinput.ThesescoresshowthatRBCdoesthebestfollowedbyNAL-RandHPF.ThedierencebetweenRBCandlineargainis6dB.ThedierencebetweenRBCandDSLforcellphonespeechisabout3dB. 42

PAGE 43

Table2-1: ThekfconstantforPOGO Freq250500100020004000 Table2-2: ThekfconstantforNAL Freq2505001000200030004000 Figure2-1. Classicationofexistinghearingaidttingmethods 43

PAGE 44

GainsprescribedbytheFig6method Figure2-3. Input-Outputcurveat2kHzobtainedfromthevisualinputoutputlocator 44

PAGE 45

DSLprescribedgainfordierenthearingloss Figure2-5. Recruitmentbasedcompensationsystem 45

PAGE 46

Computationofgainbasedonloudnessrecruitment Figure2-7. Estimateddependenceofrecruitmentrangeonhearingloss 46

PAGE 47

Compressioninput-outputandgaincurves 47

PAGE 48

b SubjectiveHINTresultsforhearing-impairedA)AverageHINTscoreswithvaryinglterbanksizeB)Averageaudiogramofthehearing-impairedlisteners 48

PAGE 49

b SubjectiveMOSresultsA)AverageMOSscoresforthehearing-impairedB)Audiogramofthehearing-impairedlisteners 49

PAGE 50

AudiogramonthephoneJavamidlet Figure2-12. Audiogramsofallthehearing-impairedlisteners 50

PAGE 51

Hearinglosssimulationsystem Figure2-14. ThePESQobjectivespeechqualityscorefornormal-hearingandhearing-impaired 51

PAGE 52

b c Spectrogramofhearing-impairedforA)TypicalmildtosevereSNHLB)Linear-AmpliedspeechC)RBCampliedspeech

PAGE 53

b c Spectrogramofnormal-hearingforA)Typicalnormal-hearingB)Linear-AmpliedspeechC)RBCampliedspeech

PAGE 54

b SubjectiveMOSresultsA)AverageMOSscoresforthehearing-impairedB)Audiogramofthehearing-impairedlisteners 54

PAGE 55

b SubjectiveMOSresultsA)AverageMOSscoresforthenormal-hearingB)Audiogramofthenormal-hearinglisteners 55

PAGE 56

Speechintelligibilityindex(SII)scoresfornormal-hearingasafunctionofSNR 56

PAGE 57

b SubjectiveHINTresultswithwidebandspeechforthehearing-impairedA)AverageHINTscoresB)Audiogramofthehearing-impairedlisteners 57

PAGE 58

b SubjectiveHINTresultswithcellphonespeechforthehearing-impairedA)AverageHINTscoresB)Audiogramofthehearing-impairedlisteners 58

PAGE 59

b SubjectiveHINTresultswithcellphonespeechforthenormal-hearingA)AverageHINTscoresB)Audiogramofthenormal-hearinglisteners 59

PAGE 60

Environmentalnoisedetrimentallyaectstheintelligibilityofspeech[ 61 ]andthiseectismorepronouncedforpeoplewithhearingimpairment.Speechisahighlyredundantsignal.Inamoderatelynoisyenvironment,anormal-hearinglistenerwillbeabletounderstandwhatisbeingsaidevenifsomepartsofthespeecharemaskedbynoisebyvirtueoftheredundantnatureofspeech.Hearing-Impairedlistenersdealwithalessredundantspeechsignalbecauseofthenatureoftheirhearingloss[ 62 ].Thisimpliesthateveniftheenvironmentalnoisemasksasmartportionoftheremainingspeech,theintelligibilitywillbedegradedsignicantly.Thecochleaanalyzessoundbymeansofagroupofhighlyoverlappingnarrowbandlters.Theseltersarecalledthecriticalbandsandplayanimportantroleinnoisemasking.Onlythenoisewhichfallswithinthesamecriticalbandasspeechcanmaskthespeech.Butthesamenoisewillmasktoalesserextent,signalsinhigherfrequencybandsbecauseofthehighlyoverlappedstructureofthecriticallterbank.Thiseectiscalledtheupwardspreadofmaskinganditincreaseswithincreaseinnoiseintensity.Thisisalsowhylowfrequencysoundsarebetterspeechmaskers.Forthehearing-impairedthecriticalbandswillbemorebroadandhencetheupwardspreadofmaskingincreases.Thisiswhyhearing-impairedlistenersrequirea9dBincreaseinSNR,whencomparedtonormal-hearinglisteners,inordertounderstandspeechinnoise[ 25 ].Thischapterwilldiscussthedevelopmentofanoiserobustrecruitmentbasedcompensation(NR-RBC)algorithm. 60

PAGE 61

3-1 providesalistofcellphonenoisesourcesandsuggestspossiblewaystoreduceit. TheRBCalgorithmusestheaudiograminquietinformationtoprescribethegains.IfmaskedthresholdsofhearingarecalculatedthentheycanbeusedintheplaceofthethresholdsinquietintheRBCestimationmethod.Thealgorithmwillthenvarythegainandcompressionbasedonboththethresholdsofhearingandtheenvironmentalnoise.Thismodiedalgorithmiscalledthenoiserobustrecruitmentbasedcompensation(NR-RBC)method. Figure 3-1 showstheblockdiagramoftheprocedure.HereS(n)istheincomingcellphonespeechsignalwhichistobeenhanced.Processingiscarriedoutinthetimedomainusingaframe-by-frameapproach.S(n)isfedtoalterbankwhichhas18bandswithcenterfrequenciesandbandwidthasshowninTable 3-2 .Theenvironmentalnoiseispickedupbythecellphone'scalibratedmicrophoneandisreferredtoasY(n).Themicrophonealsopicksuptheuser'svoice.Inordertoidentifywhichframescontainnoise,Y(n)passesthroughavoiceactivitydetectionblock.Thenoiseframesarethenfedtoanoiseestimationblockwhichprovidesanestimateofthenoiseforeachoctave-band.Thisoctave-bandnoiseestimateisusedtocomputethenoisemaskedthresholds.Thegaincomputationblockusestheenergyperband,thenoiseestimateandtheuser'shearingthresholdstoprescribeagainandcompressiontermforeachbandaspertheNR-RBC 61

PAGE 62

3-1 ).DuringpausesintheconversationY(n)picksuptheenvironmentalnoise.Usingavoiceactivitydetectionsystemtheframescanbemonitoredfornoiseandspeech.Ifanoiseagissetthenthenoisepowerestimateisthenupdated.UsingthesinglemicrophonesystemanestimateoftheenvironmentalnoiseN(w)atthelistenersendhastobecalculated.Thiswillbedoneduringpausesintheconversation.TechniqueslikeMinimaControlledRecursiveAveraging(MCRA)[ 63 ]methodandothers[ 64 ]areavailableforrobustestimationofnoise.Weusedasimplevoiceactivitydetectorbasedonspectraldistance. 65 ],[ 66 ].Asarststepacriticalbandanalysishastobecarriedoutinordertoknowwhichspeechbandsoftheincomingcellphonesignalwillbeaectedbytheenvironmentalnoise.ThiscanbeachievedbypassingthenoisethroughlterbankstructuresimilartotheoneusedinRBC 2-5 .Whilethiswillleadtoanaccurateanalysis,itwillbecomputationallyinecienttoimplementonthephone.AwaytoworkaroundthisistogrouptogethertheFFTbinsbasedonthecriticalbandcenterfrequenciesandbandwidth.Table 3-2 liststhecriticalbandnumber,centerfrequency,bandwidthandtheFFTbindetailsforabinsizeof256andasamplingfrequencyof8000Hz. CriticalbandanalysisiscarriedoutonthepowerspectrumofthesignalovertheFFTbinswhichcorrespondtoeachcriticalband.18criticalbandscoverthecellphonefrequencyrangeupto4000Hz.Sincethecriticalbandsarehighlyoverlappedstructures 62

PAGE 63

67 ]isgivenbyEquation 3{1 10log10Ti=15:81+7:5(i+0:474)17:5(1+(i+0:474)2)1=2 where,iisthecriticalbandnumberandTiisthespreadingfunction. Thenextstepinvolvesthecalculationofthenoisemaskingthreshold,givenbyEquation??,whichincludesanosettermOiwhichisspeciedinTable?? ThenoisespreadthresholdhastobeconvertedbacktotheBarkorcriticalbanddomain.Thisisdonebyrenormalization.ThebarkthresholdsarecomparedtotheindividualsabsolutethresholdsofhearingHL(w)alsointhebarkscale.ThenoisemaskingthresholdT(w)foranycriticalbandwhichhasanoisethresholdlowerthantheabsolutethresholdischangedtotheabsolutethreshold. ThegainforeachchannelGdB(w)iscalculatedbyEquationset 3{3 2,whenT(w)
PAGE 64

2-13 wasused.Atypicalmildtoseveresensorineuralhearinglossandatypicalnormal-hearingwassimulatedusingtheMatlabhearinglosssimulationblock.Theaudiogramsusedwere:[102030608090]dBHLand[51015202020]dBHL.Theunprocessedcellphonespeechwaspassedthroughthehearinglossblocktogeneratespeechwithsimulatedloss.Thenspeechpreprocessedbythevariousttingalgorithmswerepassedthroughthehearinglossblocktogeneratecompensatedspeech. AmongtheHAttingalgorithmswhichincludecompression,NR-RBChasthemaximumPESQscore.Forthenormal-hearingNR-RBChadaslightlylowerPESQscorethiscouldbedueartifacts.Inordertounderstandthisthespectrogramsofsimulatednormal-hearingspeechwerecalculated. 64

PAGE 65

3-3 showsthespectrogramsforthehearing-impairedspeech,linearlyampliedspeechandthespeechcompensatedusingtheNR-RBCmethod.Theseresultsshowthatforthehearing-impairedNR-RBChasbetterspeechqualityandintelligibilitythanwithjustalineargainwhichiswhatthecellphonesvolumecontroldoes.Figure 3-4 showsthespectrogramsforthenormal-hearingspeech,linearlyampliedspeechandthespeechcompensatedusingtheNR-RBCmethod.Theseresultsalsoshowthatforthenormal-hearingNR-RBChasbetterspeechqualityandintelligibilitythanwithjustalineargain. Figure 3-5 showstheaverageoftheMOSscoresforthehearing-impaired.Forthehearing-impaired,NR-RBChasanaverageMOSscoregreaterthan`4-Good'.Figure 3-6 showstheaverageoftheMOSscoresforthenormal-hearing.Forthenormal-hearing,NR-RBChasanaverageMOSscoregreaterthan`4-Good'. 65

PAGE 66

2 weknowthatRBCandDSLshowthebestperformance.TheperformanceofNR-RBCwascomparedtoRBCandDSLwithvaryingltersizes. Figure 3-8 showstheaveragedSNRforthe10hearing-impairedsubjects,withreferencetothebaseline(lineargain).ItisclearthatNR-RBCoutperformsRBCandDSLandhasbestperformanceatN=14(Figure 3-8 ).ThedierencebetweenNR-RBCandthebestDSLisabout20dB. Figure 3-9 showstheaveragedSNRforthe10normal-hearingsubjects,withreferencetothebaseline(lineargain).ItisclearthatNR-RBCoutperformsRBCandDSLandhasbestperformanceatN=8.ThedierencebetweenNR-RBCandthebestDSLisabout13dB. 66

PAGE 67

Sourcesofcellphonenoiseandnoise-reductionmethods CellphonenoiseWaystoreducenoise ReceiversideenvironmentnoiseTopreventitfrombeingtransmitted:BeamformingandAGCTohelpyouhearbetter:RBCalgorithm(Automaticallyadjustgain)Tohelpyouhearbetter:Occludecontra-lateralear TransmittedenvironmentnoiseToreduceitseects:Spectralsubtraction VocoderandchannelnoiseToreduceitseects:Spectralsubtraction Table3-2: CriticalbandsandFFTbins Criticalbanddetails FFTdetails Noisemaskingdetails Criticalband Centerfreq Bandwidth FFTcriticalband Osettermnumber (Hz) (Hz) range(Hz) (dB) 1 50 80 094 -172 150 100 94187 -183 250 100 187312 -194 350 100 312406 -205 450 110 406500 -216 570 120 500625 -227 700 140 625781 -238 840 150 781906 -249 1000 160 9061094 -2510 1170 190 10941281 -2511 1370 210 12811469 -2512 1600 240 14691719 -2513 1850 280 17192000 -2514 2150 320 20002312 -2515 2500 380 23122687 -2416 2900 450 26873125 -2317 3400 550 31253687 -2218 4000 700 36874000 -19 67

PAGE 68

Noiserobustrecruitmentbasedcompensation(NR-RBC)system

PAGE 69

ThePESQobjectivespeechqualityscoreforvariousHAttingalgorithms

PAGE 70

b c Spectrograminnoiseofhearing-impairedforA)TypicalmildtosevereSNHLB)Linear-AmpliedspeechC)NR-RBCampliedspeech

PAGE 71

b c Spectrograminnoiseofnormal-hearingforA)Typicalsimulatednormal-hearingB)Linear-AmpliedspeechC)NR-RBCampliedspeech

PAGE 72

b SubjectiveMOSresultsA)AverageMOSscoresforthehearing-impairedB)Audiogramofthehearing-impairedlisteners 72

PAGE 73

b SubjectiveMOSresultsA)AverageMOSscoresforthenormal-hearingB)Audiogramofthenormal-hearinglisteners 73

PAGE 74

TheSIIscoresforsimulatednormal-hearingasafunctionofSNR Figure3-8. TheHINTscoresforRBC(A1),NR-RBC(A2)andDSL(A3)withvariationofltersizeforhearing-impaired 74

PAGE 75

TheHINTscoresforRBC(A1),NR-RBC(A2)andDSL(A3)withvariationofltersizefornormal-hearing 75

PAGE 76

Theauditorycortexundergoesphysiologicalandanatomicalchangesoveraperiodoftimewhenpresentedwithalteredauditorysignalinputs.InthecaseofapersonwithmildtosevereSNHL,thealteredauditorysignalwillhavelittleornohighfrequencies.Moore[ 68 ]providedareviewofstudiesshowingevidenceofplasticityintheauditorysystemoftheadultbrain.Becauseofthisbrainplasticity,ittakestimefortheaidedhearing-impairedlistenerstofullyusethehigh-frequencyinformationthattheywerepreviouslynotusedtohearing.Thisisknownastheacclimatizationeect.Thetimeperiodforacclimatizationisdenedastheperiodbetweenwhenthehearinglosswasnoticedandwhenthehearingaidwastted. Acclimatizationismorepronouncedfornewhearingaidusersandaectsthehearingaidttingprocedure.Forarsttimehearingaiduser,theaudiologistwillrstmeasuretheamountofloss,discussthevarioushearingaidoptions(styles,binauralormonoaural)andthenchooseamakeandmodelofahearing-aid.Earmoldmeasurementsofthepatientarethenmade.Thehearingaidwillarriveafter2-3weeksandthepatientwillbe`t'withthehearingaid.Fittingistheprocedurebywhichthehearingaidparametersaretunedforthepatient'shearingloss.Usually,eachhearingaidisaccompaniedbyaCDwiththecompany'sproprietaryttingsoftwarewhichalsoallowsselectingcertainestablishedttingprocedureslikeDSLandNAL.Aslongastheinitialttingparametersdonotcauseanydiscomfort,theywillnotbemodiedduringtherstvisit.Duringthefollowupvisits,theaudiologistwillne-tunetheparametersbasedonverbalfeedbackfromthepatient.Theverbalfeedbackisdescriptiveandindicateshowcertainsoundsarenowbeingperceivedwiththehearing-aid.Thepatientisnotaskedtoratethesoundsonanyscale.Thene-tuningprocessisrepeatedovermultiplevisitsuntilthehearingaiduserissatisedwithaparticulartting.Thefollowupvisitsareusuallyapartbyacoupleofweeks. 76

PAGE 77

ThethreemainpsychoacousticphenomenaassociatedwithSNHL,elevatedthreshold,loudnessrecruitment,andfrequencyblurring,lowerthespeechintelligibilityforhearing-impairedlistenersbydegradingthespeechcues.Ourpreviousresearchinthisarea[ 69 ]hasshownthatSIbasedttingmethodologiesshowbetterperformanceinnoiseandotherrealworldscenarios.Henceitisbettertousearatingscale,asshowninTable 4-1 ,basedonSIwhere1islowintelligibilityand5ishighintelligibility.Itisassumedthatthelistenersarenottrainedlisteners,thatthespeechstimuliaresentencesandthatthelistenershavenopriorcuesaboutthespeechstimuli. 77

PAGE 78

4.2.1HearingAidFittingDataCollection 4-2 showssomeofthettingparametersprovidedbythelistedPhonakHAs.AllHAsallowfrequencybasedne-tuningofthegainparametersandthemaximumpoweroutput(MPO)valuesatoneorseveralinputsignallevels.Somehearingaidsalsoallowmodicationofthecompressionparameters(CR,TK,TKknee)acrossfrequency. Figure 4-1 showsthevariationofthe50dBand80dBgainsfortheClaroHA,asprescribedbyDSL,betweentherstandlasttting.FortheClaro,morepeopletendtopreferhavinglessergainatthenal-ttingthanattheinitial-tting.ThisgoesagainstourhypothesisthatpatientswillslowlyincreasetheHFgain.Thiscouldbeexplainedby 78

PAGE 79

15 ]. Figure 4-2 showsthevariationofthe40dBand80dBgainsfortheSaviaHA,asprescribedbyNAL-NL1,betweentherstandlasttting.FortheSaviamorepeoplepreferhavinghighergainatthenal-ttingthanattheinitial-tting.Thiscouldbeexplainedbytheacclimatizationprocess. Figures 4-3 and 4-4 showthevariationofthecompressionparametersfortheClaroandSaviaHA.ItcanbeseenthatmorepeopletendtonotchangeTKforbothClaroandSavia.ForClaroHAs,thechangeincompressionratiodoesnotfollowanyconclusivetrendwhileforSaviathecompressionratioseemstoincreaseatthenaltting.ThiscouldbeexplainedbythefactthatNAL-NL1prescribeslesserCRatallfrequenciescomparedtotheothersuprathresholdmethods[ 15 ]. InordertostudythevariationofthealltheparametersforeachHA,thetrendinchangeforeachparameterwasrstaveragedacrossthefrequency.ThemaximumtrendacrossthisaveragewasthenpickedupandplottedforeachHAinFigures 4-5 to 4-9 FortheClarowhichwastbyDSL,fromFigure 4-6 itappearsthatthegainparametersdecreaseatthenalttingwhiletheTK,MPOparametersremainthesameandtheCRdecreaseatthenaltting. FortheNAL-NL1basedHAsitappearsthatthefrequencyaveragedgainparametersincreaseatthenaltting.ForSavia(Figure 4-5 ),Extra(Figure 4-7 ),andValeo(Figure 4-8 )thefrequencyaveragedcompressionparametersremainthesameacrossthettingstages.ForEleva(Figure 4-9 )andPerseo(Figure 4-10 )thefrequencyaveragedcompressionparametersdecreaseacrossthettingstages. 70 ]andto 79

PAGE 80

71 ].J.M.Kates[ 72 ]studiedthefeasibilityofusingNNstoderiveHAprescriptiveprocedures.KatesconcludedthatthefactorswhichaectedtheaccuracyofaNNbasedttingmethodwerethetrainingdatabasesize,thevariabilityofthepatient'sresponsesandthevariabilityofthehearingloss.Gao[ 73 ]proposedanewhearingprostheticssimilartotheoneproposedbyKateswhichwasbasedonNNsandfuzzylogic. RatherthanusingNNstoreplaceexistingttingalgorithms,weproposetousetheNNsalongwiththeexistingttingalgorithmstomodeltheacclimatizationeect.TheNNwastrainedonthemulti-sessionttingdatausingsupervisedlearning.Theinputwastheinitialttingdataandthedesiredwasthenalttingdata.Inadditiontothettingdata,someHAuserspecicparameterssuchaspatient'sage,degreeofhearingloss(HL)andthenumberofyearswithHLwereusedasinputsduringtraining. Figure 4-11 showsthestructureoftheNNusedintheacclimatizationmodelingsetup.TheNNusedtomodelacclimatizationwasthemulti-layerperceptron[ 74 ].Ithad7sigmoidalinputnodesand7linearoutputnodes.Thenumberofhiddennodeswasvarieduntilthelowesterrorwasobtainedandwasfoundtobe4.TrainingwascarriedoutusingtheLevenberg-Marquardt[ 75 ]methodwithaninitialgloballearningrate(LR)of0.01.Crossvalidationwasusedtostopthetraininginordertopreventover-training.TheNNwastrainedontheSaviadatawhichhadthemostnumberofdatapoints. 80

PAGE 81

4-12 to 4-17 showtheresultsoftrainingfortheSaviaHA.Inthegures,theredcurveshowstheMSEbetweentheinitialttingvaluesandthedesiredoroptimalvalues.Fromthegures,itcanbeobservedthatthegainspredictedbytheNNareclosertotheoptimalvaluesthanbyjustusingtheinitialvalues.Thereexistssomeerrorbetweenthepredictedandtheoptimalvaluesespeciallyatthelowfrequenciesandthismightberesolvedbyincreasingthetrainingdatabase. Fromthegures,itcanbeseenthattheneuralnetworksucceedsinmodelingthetrendwithacertainamountoferror.Theerrorcanbebroughtdownbyincreasingthetrainingdatabase.TheMSEoftheoptimumsettingisalwayslessthanthatwiththeinitialsetting. Table4-1: Speechintelligibilitybasedttingsatisfactionscale SpeechintelligibilityScore Speechisneverintelligible1Speechisrarelyintelligible2Speechissometimesintelligible3Speechisusuallyintelligible4Speechisalwaysintelligible5

PAGE 82

Phonakhearingaidttingparameters HearingAidsFittingInputFrequency(kHz)NoofNoofparameterslevel(dB)parametersPatients SaviaGain40[0.30.881.72.43.57]630Gain60[0.30.881.72.43.57]630Gain80[0.30.881.72.43.57]630CR-[0.30.881.72.43.57]630TK-[0.30.881.72.43.57]630MPO-[0.30.881.72.43.57]630

PAGE 83

b ComparisonofchangefrominitialtonalstageforClaroparameterA)50dBGainB)80dBGain 83

PAGE 84

b ComparisonofchangefrominitialtonalstageforSaviaparameterA)50dBGainB)80dBGain 84

PAGE 85

b ComparisonofchangefrominitialtonalstageforClaroparameterA)TKandB)CR 85

PAGE 86

b ComparisonofchangefrominitialtonalstageforSaviaparameterA)TKandB)CR 86

PAGE 87

PhonakSaviamaximumtrendinttingparametervariationaveragedacrossfrequencies Figure4-6. PhonakClaromaximumtrendinttingparametervariationaveragedacrossfrequencies 87

PAGE 88

PhonakExtramaximumtrendinttingparametervariationaveragedacrossfrequencies 88

PAGE 89

PhonakValeomaximumtrendinttingparametervariationaveragedacrossfrequencies 89

PAGE 90

PhonakElevamaximumtrendinttingparametervariationaveragedacrossfrequencies 90

PAGE 91

PhonakPerseomaximumtrendinttingparametervariationaveragedacrossfrequencies 91

PAGE 92

StructureoftheMLPusedtomodelmulti-sessionttingtrends 92

PAGE 93

b PhonakSavianeuralnetworkmodelingresultsfor40dBgainA)testontraindataandB)testonnewdata 93

PAGE 94

b PhonakSavianeuralnetworkmodelingresultsfor60dBgainA)testontraindataandB)testonnewdata 94

PAGE 95

b PhonakSavianeuralnetworkmodelingresultsfor80dBgainA)testontraindataandB)testonnewdata 95

PAGE 96

b PhonakSavianeuralnetworkmodelingresultsforCRparameterA)testontraindataandB)testonnewdata 96

PAGE 97

b PhonakSavianeuralnetworkmodelingresultsforTKparameterA)testontraindataandB)testonnewdata 97

PAGE 98

b PhonakSavianeuralnetworkmodelingresultsforMPOparameterA)testontraindataandB)testonnewdata 98

PAGE 99

TheUScensusbureaustatesthat50millionpeople,nearlyone-fthoftheUSpopulation,areinsomewaydisabled.Amongthis28millionpeoplearehearing-impaired.Withsuitabletechnologicalassistancethesemenandwomen(notablytheagingbabyboomers)mayprolongtheirindependenceandreducetheirneedforspecializedcare.Theirqualityoflifewillbeimproved.Whilenoproductcanbedesignedsothateverysinglepersonintheworldcanuseit,theintentistomaximizethepotentialofeachdevice.Thisdissertationproposesusingthecellphoneasanassistivelisteningdevice.Thiswillenablethe20millionhearing-impairedpeoplewhodonotusehearingaidsunderstandcellphonespeechbetter.Thiswillalsohelpnormal-hearinglistenersespeciallyinnoisesituations. Sensorineuralhearingloss(SNHL)ismostlycausedbydamagetotheouterhaircells(OHC).Soinparticular,thehearinglosscompensationalgorithmhastoreplacethedamagedouterhaircells.Anovelalgorithmbasedonthefrequencyindependentrelationshipbetweenhearinglossandrecruitmentwasdeveloped.Thisrecruitmentbasedcompensation(RBC)algorithmprescribesboththegainsandcompressionparameters.RBCshowsa15dBimprovementinspeechintelligibilitywhencomparedtothebaselinealgorithm(lineargain)forthehearing-impaired.ForthenormalhearingtheSNRdierencebetweenRBCandlineargainis6dB.ByprovidingfrequencydependentgainandcompressionthedecreasedaudibilityanddecreaseddynamicrangeaspectsofSNHLareovercome.Thisiscarriedoutbyprocessingthespeechsignalin14lterbanks.ThelterbankshavetheircentersequallyspacedinmelfrequencyandissodesignedkeepingtheauditoryprocessingoftheOHCsinmind. TheOHCslosetheirabilitiestoincreasethesensitivityofthecochleaforfrequenciestowhichtheaectedpartofthecochleaistuned.Psychoacoustically,thisshowsupasattertuningcurves.Becauseofthis,noisehasagreatermaskingeectfor 99

PAGE 100

Theauditorycortexundergoesphysiologicalandanatomicalchangesinthepresenceofalteredauditoryinput.Becauseofthisbrainplasticity,ittakessometimeforthehearing-impairedtolearntofullyusethehigh-frequencyinformationthattheywerepreviouslynotusedtohearing.Thisisknownastheacclimatizationeect.Multiple-sessiondataforanumberofPhonakhearingaidswascollectedandanalyzed.Neuralnetworkwereusedtomodeltheacclimatizationeectinhearingaidtting.Lowmeansquareerror(MSE)fortestonnewdatawasobtained. Threenovelalgorithmsallbasedontherationaleofmaximizingspeechintelligibilitywerecreated.RBCisaimedtowardshelpingbothnormal-hearingandunaidedhearing-impairedlistenersunderstandcellphonespeechbetter.NR-RBCisanoiserobusttechniquewhichenhancesthehearinginnoise.Acclimatizationmodelingisproposedtoimprovethequalityoftheinitialtfortheaidedhearing-impairedlistener.Thehearingenhancementalgorithmsweretestedextensivelyintermsofobjectiveandsubjectivemeasuresofspeechquality(SQ)andintelligibility(SI)onbothnormal-hearingandhearing-impairedsubjects 100

PAGE 101

101

PAGE 102

Thischapterdiscussestheresultsofthecellphonehearingevaluationsurvey.84hearing-impairedparticipantsanswered20hearingaidandcellphonerelatedquestionsandprovidedotherrelevantdemographicinformation.ThequestionnaireusedinthesurveyisavailableinAppendix B A-1 .Hearingaidusageexperiencevariedamongthesubjects.Subjectsrangedfrom23to89yearsofagewithameanageof66.32yearsold.Fifty-twomalesand32femalesparticipatedinthesurvey. 102

PAGE 103

A-2 indicatestheself-reportedunderstandingofcellphonespeechinquietandinnoiseforbothaidedandunaidedhearing-impairedparticipants. 103

PAGE 104

6 ]indicateshowhearingaidandcellphonemanufacturersshouldmeasuretheEMinterference.ThemeasurementsaretranslatedintoM-ratingsinwhichthehigherratingsindicatealowerlikelihoodofinterference.HandsetsthatreceiveahearingaidcompatibilityratingofM3orhigherhavemetorsurpassedFCCrequirement.TheFCChasrequiredthatcellphonecompanieshave50%oftheirhandsetsmeetaminimumANSIratingofM3orbetterbyFebruary18,2008. FigureA-1. Degreeofhearingimpairmentamongsurveyparticipants 104

PAGE 105

Degreeofhearingimpairmentforsurveyparticipants 105

PAGE 106

1. Isyourhearingloss: (a) Mild-Alittledicultyhearingspeech (b) Moderate-Moredicultyhearingspeech (c) Severe-Alotofdicultyhearingspeech (d) Profound-Sobadthathearingaidsmaynothelp 2. DoyouuseHearingaids? (a) Yes (b) No 3. YourLeftEarHearingaidis: (a) None (b) ITE-InTheEar (c) BTE-BehindtheEar (d) CIC-CompletelyInCanal (e) Other 4. YourRightEarHearingaidis: (a) None (b) ITE-InTheEar (c) BTE-BehindtheEar (d) CIC-CompletelyInCanal (e) Other 5. WhatisthemakeandmodelofyourHearingaid? 6. Checkwhicheveristrue: (a) Icanunderstandspeechoverthetelephonewithmyhearingaid 106

PAGE 107

Icanunderstandspeechoverthetelephoneinnoisyenvironmentswithmyhearingaid (c) Icanunderstandspeechoverthetelephonewithoutmyhearingaid 7. DoesyourHearingaidhaveatelecoil? (Afeatureavailableonmanyhearingaidsisthetelecoilort-switchort-coilwhichaidsinhearingtelephoneconversations.) (a) Yes (b) No 8. IfyourHearingaidhasatelecoil,doyouuseyourtelecoilwithyourcellphone? (a) Yes (b) No (c) Idon'tusecellphones 9. Doyoufrequentlyusecellphones? (a) Yes (b) No 10. Checkwhicheveristrue:Whathasbeenyourgeneralexperiencewithcellphones? (a) Icanunderstandspeechonthecellphone (b) Icanunderstandspeechonthecellphonesinnoisyenvironments (c) Ihavetroublehearingmycellphonering (d) Icannotunderstandspeechonthecellphonesbecause: Whatisthemakeandmodelofthecellphoneyouuse? (ExampleMake=MotorolaModel=v300) 12. IsitaFlipphone? (a) Yes 107

PAGE 108

No 13. Doyouuseaneckloopforcellphoneconversations? (Aneckloopisanecklace-sizeloopofwirewornaroundtheneckofsomeonewhohasahearingaidwithatelecoil.) (a) Yes (b) No 14. WhichCellphoneNetworkproviderdoyouuse? (a) Cingular (b) Tmobile (c) Verizon (d) Sprint (e) Other 15. Howmanyminutespermonthdoyoutalkonthecellphone? (a) Justforemergencies (b) (c) 200-400 (d) 400+ (e) Don'tknow 16. Checkwhicheveristrue: (a) Iwishmycellphonewouldbelouder (b) MycellphonecreatenoiseinmyHearingaid (c) Thisnoisepreventsmefromusingmycellphone (d) ApersontalkingnearmeonacellphoneproducesnoiseinmyHearingaid 17. DoesyourcellphonebacklightcausenoiseinyourHearingaid? (a) Yes (b) No (c) Notnoticed 108

PAGE 109

WouldyoubeinterestedinacombinationHearingaidandcellphone? (a) Yes (b) No 19. Isthereenoughinformationavailableonlinetohelpyouchoosethecellphonerightforyou? (a) Yes (b) No 20. Whatinformationregardingcellphonesandhearingaidswouldyouliketoseeavailable? 109

PAGE 110

Inordertobetterunderstandtheneedsofthehearing-impairedpopulationtwofocusgroupswithhearing-impairedparticipantswereconducted.Thischapterprovidesasynopsisofthemainthemeswhichwereobservedatthetwofocusgroup. C.2.1AidedCellPhoneListeningProblems 110

PAGE 111

111

PAGE 112

Thehumanear,theorganofhearingandbalance,isthebestexampleofanengineeringmasterpiece.Itenablesustohearsoundsrangingfrom20Hz-20kHzwithadynamicrangeof0-130dB.Anatomically,thehumanearcanbedividedintothreeparts:theouterear(pinna,auditorycanal),themiddleear(ossicles,eardrum,ovalwindow)andtheinnerear(cochlea,semicircularcanals).Figure D-1 ,showstheinternalstructureoftheear. Thehumanpinnaissymmetric,pointsforwardandhasacurvedstructure.Itfocusessoundpressurewavesintotheauditorycanal.Thestructureofthepinnaaidsinsoundlocalization.Horizontallocalizationismadepossiblebecauseofinter-auraltimeandintensitydierenceswhileverticallocalizationismadepossiblebecauseofthefrequencyshapingofthesoundbythecurvesofthepinna.Theauditorycanalwhichisaround2.7cminlengthactsasa1 4waveclosedtuberesonatorandbooststhe2-5kHzregionby15dB.Thebroadresonancepeakisbecausetheclosedendoftheauditorycanalistheplianteardrumortympanicmembrane. Themiddleearconsistsoftheeardrum,theossicles(malleus,incusandstapes)andtheovalwindow.Theossiclestranslatethesoundpressurewavetovibrationsinthecochlea.Theyprovideimpedancematchingsincetheacousticimpedanceoftheuidinthecochleaisabout4000timesthatofair.Theossiclesprovideamplicationbyleveraction(3x)andbytermsofareaamplication(15x).Theossiclesalsohelpblockveryloudlowfrequencysoundsbymeansofthestapediusreex.Thestapestransmitsvibrationstotheovalwindowontheoutsideofthecochlea.Thismovestheuidinthecochleawhichformsatravelingwave,withapeakatonelocationalongthelengthofthecochlea.Conductivehearinglossoccurswhensoundisnotconductedecientlytothecochleathroughtheossicles.Thecochleaisintactforconductivehearingloss. 112

PAGE 113

D-2 ).Threerowsconsistofouterhaircells(OHCs)andonerowconsistsofinnerhaircells(IHCs).Eachhaircellhashundredsoftinystereocilia.ThestereociliaoftheOHCsareembeddedinthetectorialmembrane.ThetravelingwavebendstheIHC'sstereociliaandthisproducesactionpotentials. TheaerentIHCstransmitsignalstothebrainviatheauditorynerve.TheeerentOHCsreceiveneuralinputfromthebrainwhichinuencesitsmotilityaspartofthecochlea'smechanicalpre-amplier.TheOHCshelptheIHCssensesoftsoundsbysharpeningthepeakofthetravelingwave.Basedonfeedbackfromthebrain,theOHCsmechanicallyshrinkpullingdownthetectorialmembrane.OncetheIHCsciliabrushesagainstthetectorialmembrane,actionpotentialsaregenerated.Inadditiontoamplifyingsoftsounds,OHCsalsosharpenthepeakofthetravelingwaveresultinginhighfrequencyresolution.Sensorineuralhearinglossoccursbecauseofdamagetothehaircells.Hearinglossduetoagingorpresbycusisisatypeofsensorineuralhearinglossandoccursduetowearandtearofthehaircells.Ahearinglossupto60dBHLcanbeconsideredtobebecauseofOHCsandanythinghigherthan80dBHLisbecauseofbothIHCandOHCdamage.Figure D-3 ,showsthehaircellsforapersonwithnormal-hearingandapersonwithseverehearingloss.DamagetotheOHCsindicatesalowerfrequencyresolutionandtheinabilitytohearsoftsoundswhiledamagetotheIHCsindicatesthatthesoundinformationisnotbeingsenttothebrain. ThereisatonotopicmappingalongthelengthoftheBM(Figure D-4 ).EachpartoftheBMhasacharacteristicfrequencyofmaximumvibrationwhichdependsonitsrelativeposition.Atthebaseofthecochlea(neartheovalwindow),theBMisstiandthinand 113

PAGE 114

FigureD-1. Structureofthehumanear[ 76 ] 114

PAGE 115

HaircellsoftheOrganofcorti[ 77 ] b ElectronmicrographoftheorganofcortiforA)Normal-hearingB)Severehearingloss[ 78 ] 115

PAGE 116

Frequencysensitivityofthebasilarmembrane 116

PAGE 117

[1] S.Arlinger,\Negativeconsequencesofuncorrectedhearingloss{areview,"Interna-tionalJournalofAudiology,vol.42,pp.S17{20,2003. [2] \CTIAsemi-annualwirelessindustrysurvey,"(updated2007;citedDecember2007).[Online].Available: http://les.ctia.org/pdf/CTIA Survey Mid Year 2007.pdf ,CTIA,Washington,DC. [3] Statisticsabouthearingdisorders,earinfections,anddeafness.(updatedJanuary2007;citedDecember2007).[Online].Available: http://www.nidcd.nih.gov/health/statistics/hearing.asp .NIDCD.Bethesda,MD. [4] M.C.Killion,\TheSINreport:Circuitshaven'tsolvedthehearing-in-noiseproblem,"TheHearingJournal,vol.50,p.10,1997. [5] M.Skopec,\Hearingaidelectromagneticinterferencefromdigitalwirelesstelephones,"IEEETransactionsOnRehabilitationEngineering,vol.6,pp.235{239,1998. [6] [7] J.Rodman,\Theeectofbandwidthonspeechintelligibility,"WhitePaper,Polycom,2003. [8] B.C.J.Moore,\Perceptualconsequencesofcochlearhearinglossandtheirimplicationsforthedesignofhearingaids,"EarandHearing,vol.17,pp.133{161,1996. [9] T.Venema,CompressionforClinicians.SanDiego,CA:SingularPublishingGroup,1999. [10] Noiseinducedhearinglossfacts.(updated2007;citedDecember2007).[Online].Available: http://www.hei.org/news/facts/nihlfact.htm .HEI.LosAngeles,CA. [11] C.Portnu.SafelisteninglevelsforappleiPod.(updatedOctober2006;citedDecember2007).[Online].Available: http://www.physorg.com/news80304823.html .Boulder,CO. [12] [13] [14] J.W.HallandH.G.Mueller,Audiologists'DeskReferenceVolumeI:DiagnosticAudiologyPrinciplesProceduresandProtocols.SanDiego,CA:SingularPublishingGroup,1996. 117

PAGE 118

H.Dillon,HearingAids.NewYork:ThiemeMedicalPublishers;1stedition,2001. [16] M.C.Killion,\SNRloss:Icanhearwhatpeoplesay,butican'tunderstandthem,"TheHearingReview,vol.4,pp.8{14,1997. [17] J.H.MacraeandH.Dillon,\Gain,frequencyresponse,andmaximumoutputrequirementsforhearingaids,"JournalofRehabilitationResearchandDevelopment,vol.33,no.4,pp.363{76,1996. [18] J.C.SteinbergandM.B.Gardner,\Thedependenceofhearingimpairmentonsoundintensity,"JournaloftheAcousticalSocietyofAmerica,vol.9,no.1,pp.11{23,1937. [19] B.A.Henry,C.W.Turner,andA.Behrens,\Spectralpeakresolutionandspeechrecognitioninquiet:Normalhearing,hearingimpaired,andcochlearimplantlisteners,"JournaloftheAcousticalSocietyofAmerica,vol.118,pp.1111{1121,2005. [20] B.R.GlasbergandB.C.J.Moore,\Psychoacousticabilitiesofsubjectswithunilateralandbilateralcochlearhearingimpairmentsandtheirrelationshiptotheabilitytounderstandspeech,"ScandinavianAudiology,vol.32,pp.1{25,1989. [21] E.ZwickerandK.Schorn,\Psychoacousticaltuningcurvesinaudiology,"Audiology,vol.17,pp.120{40,1978. [22] E.M.DanaherandJ.M.Pickett,\Somemaskingeectsproducedbylow-frequencyvowelformantsinpersonswithsensorineuralhearingloss,"JournalofSpeechandHearingResearch,vol.18,pp.261{71,1975. [23] J.R.DubnoandA.B.Schaefer,\Frequencyselectivityforhearing-impairedandbroadband-noise-maskednormallisteners,"QuarterlyJournalofExperimentalPsychology,vol.43,pp.543{64,1991. [24] E.M.Danaher,M.P.Wilson,andJ.M.Pickett,\Backwardandforwardmaskinginlistenerswithseveresensorineuralhearingloss,"Audiology,vol.17,pp.324{38,1978. [25] S.Hygge,J.Ronnberg,andB.Larsby,\Normal-hearingandhearing-impairedsubjects'abilitytojustfollowconversationincompetingspeech,reversedspeech,andnoisebackgrounds,"JournalofSpeechandHearingResearch,vol.35,pp.208{15,1992. [26] B.C.J.MooreandB.R.Glasberg,\Simulationoftheeectsofloudnessrecruitmentandthresholdelevationontheintelligibilityofspeechinquietandinabackgroundofspeech,"JournaloftheAcousticalSocietyofAmerica,vol.94,pp.2050{2062,Oct.1993. [27] Y.NejimeandB.C.J.Moore,\Simulationoftheeectofthresholdelevationandloudnessrecruitmentcombinedwithreducedfrequencyselectivityontheintelligibilityofspeechinnoise,"JournaloftheAcousticalSocietyofAmerica,vol.102,no.1,pp.603{615,1997. 118

PAGE 119

P.DuchnowskiandP.M.Zurek,\Villchurrevisited:Anotherlookatautomaticgaincontrolsimulationofrecruitinghearingloss,"JournaloftheAcousticalSocietyofAmerica,vol.98,no.6,pp.3170{3181,1995. [29] P.W.Barnett,\Overviewofspeechintelligibility,"inProceedingsoftheInstituteofAcoustics,vol.21,1999. [30] K.Kasturi,P.Loizou,M.Dorman,andT.Spahr,\Theintelligibilityofspeechwithholesinthespectrum,"JournaloftheAcousticalSocietyofAmerica,vol.112,2002. [31] N.R.FrenchandJ.C.Steinberg,\Factorsgoverningtheintelligibilityofspeechsounds,"JournaloftheAcousticalSocietyofAmerica,vol.19,p.90,1947. [32] J.D.Miller,\Eectsofnoiseonpeople,"JournaloftheAcousticalSocietyofAmer-ica,vol.56,pp.724{764,1974. [33] S.Fidell,R.Horonje,andS.Teeteller,\Eectivemaskingbandwidthsatlowfrequencies,"JournaloftheAcousticalSocietyofAmerica,vol.73,pp.628{38,1983. [34] G.ParikhandP.Loizou,\Theinuenceofnoiseonvowelandconsonantcues,"JournaloftheAcousticalSocietyofAmerica,vol.118,pp.3874{3888,2005. [35] L.M.JenstadandP.E.Souza,\Quantifyingtheeectofcompressionhearingaidreleasetimeonspeechacousticsandintelligibility,"JournalofSpeech,LanguageandHearingResearch,vol.48,pp.651{67,2005. [36] R.V.Shannon,F.Zeng,andV.Kamath,\Speechrecognitionwithprimarilytemporalcues,"Science,vol.270,pp.303{304,1995. [37] Hearinginnoisetest.(updatedJune2005;citedDecember2007).[Online].Available: http://www.californiaearinstitute.com/audiology/hint.php .CaliforniaEarInstitute.PaloAlto,CA. [38] [39] [40] MOSresultsonDVSIAMBEVocoders.(updated2006;citedDecember2007).[Online].Available: http://www.dvsinc.com/papers/toll.htm .DVSI.Westford,MA. [41] Babyboomerhearinglossstudy.(updated2006;citedDecember2007).[Online].Available: http://www.clarityproducts.com/boomer/ .ClarityandtheEARfoundation.Chattanooga,TN. [42] E.ZwickerandH.Fastl,Psychoacoustics,factsandmodels.NewYork:SpringerVerlag,1990. 119

PAGE 120

K.W.Berger,E.N.Hagberg,andD.M.Varavvas,\Comparisonofhearingthresholdlevelandmostcomfortableloudnesslevelinhearingaidprescription,"EarandHearing,vol.3,pp.30{3,1982. [44] G.KeidserandF.Grant,\Comparingloudnessnormalization(IHAFF)withspeechintelligibilitymaximization(NAL-NL1)whenimplementedinatwo-channeldevice,"EarandHearing,vol.22,pp.501{515,2002. [45] C.V.Pavlovic,\Derivationofprimaryparametersandprocedeuresforuseinspeechintelligibilitypredictions,"JournaloftheAcousticalSocietyofAmerica,vol.82,pp.413{422,1987. [46] L.E.Humes,\Evolutionofprescriptivettingapproaches,"AmericanJournalofAudiology,vol.5,pp.19{23,1995. [47] \Prescriptionofgain/output(POGO)forhearingaids,"HearingInstruments,vol.34,p.16,1983. [48] D.ByrneandH.Dillon,\Thenationalacousticlaboratories'(nal)newprocedureforselectingthegainandfrequencyresponseofahearingaid,"EarandHearing,vol.7,pp.257{265,1986. [49] M.C.Killion,\Loudness-databasisforFIG6hearing-aidttingtargets,"JournaloftheAcousticalSocietyofAmerica,vol.98,p.2927,1995. [50] R.M.Cox,\Ahands-ondiscussionoftheIHAFFapproach,"TheHearingJournal,1995. [51] R.C.Seewald,M.Ross,andM.K.Spiro,\Selectingamplicationcharacteristicsforyounghearing-impairedchildren,"EarandHearing,vol.6,pp.48{53,1985. [52] R.SeewaldandL.Cornelisse,\AsoftwareimplementationofthedesiredsensationlevelDSL[i/o]methodforttinglineargainandwidedynamicrangecompression,"Usersmanual,HealthcareResearchUnit,UniversityofWesternOntario,London,Ontario,Canada,1997. [53] E.Villvhur,\Signalprocessingtoimprovespeechintelligibilityinperceptivedeafness,"JournaloftheAcousticalSocietyofAmerica,vol.53,no.6,pp.1646{1657,1973. [54] C.S.HallpikeandJ.D.Hood,\Observationsupontheneurologicalmechanismoftheloudnessrecruitmentphenomenon,"ActaOtolaryngol,vol.50,pp.472{86,1959. [55] F.Miskolczy-Fodor,\Relationbetweenloudnessanddurationoftonalpulses.iii.responseincasesofabnormalloudnessfunction,"JournaloftheAcousticalSocietyofAmerica,vol.32,pp.486{492,1960. 120

PAGE 121

A.J.OxenhamandS.P.Bacon,\Cochlearcompression:perceptualmeasuresandimplicationsfornormalandimpairedhearing,"EarandHearing,vol.24,no.5,pp.352{366,2003. [57] C.Grasso,D.Quaglia,andL.Farinetti,\Wide-bandcompensationofpresbycusis,"inIASTEDInternationalConferenceonSignalProcessing,PatternRecognition,andApplications,Greece,2003,pp.104{108. [58] B.C.J.Moore.AnewCDfordiagnosisofdeadregionsinthecochlea-theTEN(HL)CD.(updated2004;citedDecember2007).[Online].Available: http://hearing.psychol.cam.ac.uk/dead/TENCD.html .DepartmentofExperimentalPsychology.Cambridge,England. [59] Q.Yang,\Hearingenhancementalgorithmimplementation,"UniversityofFlorida,MotorolaInternalDocument,December2006. [60] R.J.NiederjohnandD.G.Mliner,\Theeectsofhigh-passandoflow-passlteringupontheintelligibilityofspeechinwhitenoise,"JournalofAudiologyResearch,vol.22,pp.189{99,1982. [61] J.CooperandB.Cutts,\Speechdiscriminationinnoise,"JournalofSpeechandHearingResearch,vol.14,pp.332{337,1971. [62] M.A..PickenyandN.I..Durlach,\Speakingclearlyforthehardofhearingii:acousticcharacteristicsofclearandconversationalspeech,"JournalofSpeechandHearingResearch,vol.29,pp.434{445,1986. [63] I.Cohen,\Noisespectrumestimationinadverseenvironments:Improvedminimacontrolledrecursiveaveraging,"IEEETransactionsonSpeechandAudioProcessing",year="2002",volume="11",pages=466-475,. [64] K.Maleh,A.Samouelian,andP.Kabal,\Frame-levelnoiseclassicationinmobileenvironments,"inProceedingsofIEEEConferenceonAcoustics,SpeechandSignalProcessing,Phoenix,AZ,1999,pp.237{240. [65] N.Virag,\Singlechannelspeechenhancementbasedonmaskingpropertiesofthehumanauditorysystem,"IeeeTransactionsOnSpeechAndAudioProcessing,vol.7,no.2,pp.126{137,1999. [66] J.D.Johnston,\Transformcodingofaudiosignalsusingperceptualnoisecriteria,"IEEEJournalonSelectedAreasinCommunications,,vol.6,pp.314{323,1988. [67] M.R.SchroederandB.S.Atal,\Optimizingdigitalspeechcodersbyexploitingmaskingpropertiesofthehumanear,"JournaloftheAcousticalSocietyofAmerica,vol.66,p.16471652,1979. [68] C.I.Moore,M.B.Browning,andG.M.Rose,\Hippocampalplasticityinducedbyprimedburst,butnotlong-termpotentiation,stimulationisimpairedinareaCA1ofagedscher-344rats,"Hippocampus,vol.3,pp.57{66,1993. 121

PAGE 122

M.Ramani,J.G.Harris,andA.Holmes,\Comparisonofhearinglosscompensationalgorithmsusingspeechintelligibilitymeasures,"inJournaloftheAcousticalSocietyofAmerica,vol.117,Vancouver,CA,April2005,p.2604. [70] F.Feldbusch,\Identicationofnoisesbyneuralnetsforapplicationinhearingaids,"inProceedingsoftheSecondInternationalICSCSymposiumonNeuralComputation,Berlin,May2000,pp.505{510. [71] O.Arnsten,H.Koren,andT.Strom,\Hearing-aidpre-selectionthroughaneuralnetwork,"ScandinavianAudiology,vol.25,pp.259{262,1996. [72] J.M.Kates,\Onthefeasibilityofusingneuralnetstoderivehearing-aidprescriptiveprocedures,"JournaloftheAcousticalSocietyofAmerica,vol.98,no.1,pp.172{180,1995. [73] R.Gao,Y.Liu,andS.Basseas,\Nextgenerationhearingaiddevices,"inProceedingsofthe11thIEEEInternationalConferenceonToolsWithArticialIntelligence,vol.117,Chicago,IL,April1999,pp.327{331. [74] D.W.Ruck,\Themultilayerperceptronasanapproximationtoabayesoptimaldiscriminantfunction,"IEEETransactionsonNeuralNetworks,vol.1,no.4,pp.296{298,1995. [75] I.A.Manolis,\Abriefdescriptionofthelevenberg-marquardtalgorithm,"(updatedFebruary2005;citedDecember2007).[Online].Available: http://www.ics.forth.gr/lourakis/levmar/levmar.pdf ,InstituteofComputerScience,Crete,Greece. [76] D.Packard.Humanear.(Updated2006;CitedDecember2007).[Online].Available: http://commons.wikimedia.org/wiki/Image:HumanEar.jpg .Stanford,CA. [77] R.FettiplaceandC.Hackney,\Thesensoryandmotorrolesofauditoryhaircells,"NatureReviewsNeuroscience,vol.7,pp.19{29,2006. [78] A.S.Lab,\Electronmicrographsofnormalanddamagedcochlea,"(updated2005;citedDecember2007).[Online].Available: http://www.sickkids.ca/AuditoryScienceLab/section.asp?s=Hearing+Loss&sID=3243 ,UniversityofToronto,Toronto,Canada. 122

PAGE 123

MeenaRamaniwasborninthesmalltownofPalakkad,KeralainIndia,onJune1,1981.ShereceivedherbachelorsdegreeinelectronicsandcommunicationengineeringfromKumaragurucollegeoftechnologyin2002.ShereceivedhermastersdegreeinelectricalandcomputerengineeringfromtheUniversityofFloridain2004.ShejoinedtheComputationalNeuroEngineeringLaboratory(CNEL)attheUniversityofFloridain2003andhassincebeenworkingundertheguidanceofDr.JohnG.Harris.HerPhDresearchhasbeenfundedbyMotorola.Herresearchinterestsincludespeechandhearingenhancement,noisereduction,spikebasedcomputationandcochlearimplants.Duringthecourseofherstudies,MeenahasworkedasasoftwaredevelopmentandtestinginternatMicrosoft,WAandasaresearchscientistwithSoundID,CA.AftergraduationMeenawilljoinSoundIDfulltimeasaresearchscientist. 123