IEEE TRANSACTIONS ON AUDIO SPEECH AND LANGUAGE PROCES - PDF document

MURTYANDYEGNANARAYANA:EPOCHEXTRACTIONFROMSPEECHSIGNALSastheinstantofsigniÞcantexcitation.Somemethodsalsoex-ploittheperiodicitypropertyofthesignalintheadjacentcy-clesforepochextraction.Themethodproposedinthispaperassumesandexploitstheimpulse-likecharacteristicoftheexci-tation.Theintervalsbetweentheadjacentimpulsesarenotnec-essarilyequal,i.e.,theglottalcyclesneednotbeperiodiceveninshortintervalsofafew(2Ð4)glottalcycles.TheÞrstcontributiontothedetectionofepochswasduetoSobakin[10].AslightlymodiÞedversionwasproposedbyStrube[11].InStrubeÕswork,somepredictormethodsbasedonLPanalysisforthedeterminationoftheepochswerereviewed.Thesemethodsdonotalwaysyieldreliableresults.SobakinÕsmethodusingthedeterminantoftheautocovariancematrixwasexaminedcritically,anditwasshownthatthedeterminantwasmaximumifthebeginningoftheinterval,onwhichtheautocovariancematrixwascomputed,coincidedwiththeglottalIn[12],amethodbasedonthecompositesignaldecomposi-tionwasproposedforepochextractionofvoicedspeech.Asu-perpositionofnearlyidenticalwaveformswasreferredtoasacompositesignal.TheepochÞlterproposedinthiswork,com-putestheHilbertenvelopeofthehighpassÞlteredcompositesignaltolocatetheepochinstants.Itwasshownthattheinstantsofexcitationofthevocal-tractcouldbeidentiÞedpreciselyevenforcontinuousspeech.However,thismethodissuitableforan-alyzingonlycleanspeech.TheerrorsignalobtainedintheLPanalysis,referredtoastheLPresidual,isknowntocontaininformationpertainingtoepochs.AlargevalueoftheLPresidualwithinapitchperiodissupposedtoindicatetheepochlocation[13].However,epochidentiÞcationdirectlyfromtheLPresidualisnotrecommended[11],becausetheLPresidualcontainspeaksofrandompolarityaroundtheepochs.ThismakesunambiguousidentiÞcationoftheepochsfromtheLPresidualdifÞcult.AdetailedstudywasmadeonthedeterminationoftheepochsfromtheLPresidual[14],andamethodforunambiguousidentiÞcationofepochsfromtheLPresidualwasproposed.Aleast-squaresapproachforglottalinverseÞlteringfromtheacousticspeechwaveformwasproposedin[15].Inthispaper,covarianceanalysiswasdiscussedforaccuratelyperformingtheglottalinverseÞlteringfromtheacousticspeechwaveform.Amethodbasedonmaximum-likelihoodtheoryforepochdeterminationwasproposedin[16].Inthismethod,thespeechsignalwasprocessedtogetthemaximum-likelihoodepochdetection(MLED)signal.ThestrongestpositivepulseintheMLEDsignalindicatestheepochlocationwithinapitchperiod.However,theMLEDsignalcreatesnotonlyastrongandsharpepochpulse,butalsoasetofweakerpulseswhichrepresentthesuboptimalepochcandidateswithinapitchperiod.Hence,aselectionfunctionwasderivedusingthespeechsignalanditsHilberttransform,whichemphasizedthecontrastbetweentheepochandthesuboptimalpulses.UsingtheMLEDsignalandtheselectionsignalwithappropriatethreshold,theepochsweredetected.Thelimitationofthismethodisthechoiceofwindowforderivingtheselectionfunction,andalsotheuseofthresholdfordecidingtheepochs.AFrobeniusnormapproachfordetectingtheepochswasproposedin[17].Inthispaper,anewapproachbasedonsingularvaluedecomposition(SVD)wasproposed.TheSVDmethodamountstocalculatingtheFrobeniusnormsofsignalmatrices,andistherefore,computationallyefÞcient.Themethodwasshowntobeworkingonlyforvowelsegments.NoattemptwasmadeindetectingepochsindifÞcultcaseslikenasals,voicedconsonants,andsemivowels.Amethodfordetectingtheepochsinaspeechsignalusingthepropertiesofminimumphasesignalsandgroup-delayfunc-tionwasproposedin[18].Themethodisbasedonthefactthattheaveragevalueofthegroup-delayfunctionofasignalwithinananalysisframecorrespondstothelocationofthesigniÞcantexcitation.Animprovedmethodbasedonthecomputationofthegroup-delayfunctiondirectlyfromthespeechsignalwasproposedin[19].Robustnessofthegroup-delaybasedmethodagainstadditivenoiseandchanneldistortionswasstudiedin[20].Fourmeasuresofgroup-delay(averagegroup-delay,zerofrequencygroup-delay,energy-weightedgroup-delay,anden-ergy-weightedphase)andtheiruseforepochdetectionwasin-vestigatedin[21].Inthispaper,theeffectofthelengthofanal-ysiswindow,thetradeoffbetweenthedetectionrateandthetimingerror,andthecomputationalcostofevaluatingthemea-sureswereexaminedindetail.Inthispaper,itwasshownthattheenergy-weightedmeasuresperformedbetterthantheothertwomeasures.Adynamicprogrammingprojectedphase-slopealgorithm(DYPSA)forautomaticestimationofglottalclosureinstantsinvoicedspeechwaspresentedin[22]and[23].Inthismethod,thecandidatesforGCIwereobtainedfromthezerocrossingsofthephase-slopefunctionderivedfromtheen-ergy-weightedgroup-delay,andwerereÞnedbyemployingadynamicprogrammingalgorithm.Inthispaper,itwasshownthatDYPSAperformedbetterthantheexistingmethods.Epochisaninstantproperty,but,inmostofthemethodsdiscussedabove(exceptthegroup-delaybasedmethods),theepochsaredetectedbyemployingblockprocessingap-proaches,whichresultinambiguityaboutthepreciselocationoftheepochs.MostoftheexistingmethodsrelyontheLPresidualsignalderivedbyinverseÞlteringthespeechsignal.Thoughthesemethodsworkwellinmostcases,theyneedtodealwiththefollowingissues:1)selectionofparameters(orderofLPanalysis,lengthofthewindow)forderivingtheerrorsignal;2)dependenceofthesemethodsontheenergyoftheerrorsignal,whichinturndependsontheenergyofthesignal;3)theaccuracywithwhichtheepochscanberesolveddecreasesasaresultofblockprocessing;4)settingathresholdvaluetotakeanunambiguousdecisiononthepresenceofanepoch;5)thoughsomeofthesemethodsexploitperiodicityforaccurateestimationofepochlocations,theexcitationim-pulsesneednotbeperiodic.Ingeneral,itisdifÞculttodetecttheepochsinthecaseoflowvoicedconsonants,nasalsandsemivowels,breathyvoices,andfemalespeakers.Inthispaper,weproposeanewmethodforepochextractionwhichisbasedontheassumptionthatthemajorsourceofex-citationofthevocal-tractsystemisduetoasequenceofim-pulse-likeeventsintheglottalvibration.Theimpulseexcita-tiontothesystemresultsinadiscontinuityinfrequencyintheoutputsignal.Weproposeanovelapproachtodetecttheloca-tionofthediscontinuityinfrequencyintheoutputsignalbyconÞningtheanalysisaroundasinglefrequency.InSectionII,wediscussthebasicprincipleoftheproposedmethodandil-lustratetheprincipleforseveralcasesofsyntheticexcitationsignals.InSectionIII,wediscusstheissuesinvolvedinap-plyingthemethoddirectlyonspeechdata.InSectionIV,we 1604IEEETRANSACTIONSONAUDIO,SPEECH,ANDLANGUAGEPROCESSING,VOL.16,NO.8,NOVEMBER2008presentourproposedmethodtoextractepochsfromthespeechsignal.InSectionV,theperformanceoftheproposedmethodintermsofidentiÞcationaccuracyisgiven,andtheresultsarecomparedwiththreeexistingmethodsforepochextraction.InSectionVI,theperformanceoftheproposedmethodiseval-uatedfordifferenttypesofdegradations,andtheresultsarecomparedwiththeexistingmethods.Finally,inSectionVIIwesummarizethecontributionsofthispaper,anddiscusssomelim-itationsoftheproposedmethodwhichpromptfurtherinvestiga-tionforextractingepochsfromspeechsignalsrecordedinprac-ticalenvironments.II.BASISFORTHEROPOSEDETHODOFXTRACTIONSpeechisproducedbyexcitingthetime-varyingvocal-tractsystembyoneormoreofthefollowingthreetypesofexcitation:1)glottalvibration;2)frication;3)burst.Theprimarymodeofexcitationisduetoglottalvibration.Whileexcitationispresentthroughouttheproductionprocess,itisconsideredsigniÞcant(especiallyduringglottalvibration)onlywhenthereislargeen-ergyinshort-timeinterval,i.e.,whenitisimpulse-like.Theseimpulse-likecharacteristicsareusuallyexhibitedaroundthein-stantsofglottalclosureduringeachglottalcycle.Thepresenceoftheseimpulse-likecharacteristicssuggeststhattheexcitationcanbeapproximatedasasequenceofimpulses.Thisassump-tionontheexcitationofthevocal-tractsystemsuggestsanewapproachforprocessingthespeechsignalasdiscussedinthisAllphysicalsystemsareinertialinnature.Theinertialsys-temsrespondwhenexcitedbyanexternalsource.Theexcita-tiontoaninertialsystemcanbeanyofthefollowingfourtypes.ExcitationimpulseisnotintheobservedintervalofthesignalÑSinusoidalgenerator:Outputsignalisthere-sponseofapassiveinertialsystemforanimpulse,andtheimpulsesthemselvesarenotpresentintheobservedintervalsofthesignal.Sinusoidalexcitation:Sinusoidalexcitationcanbeviewedasimpulseexcitationinthefrequencydomain.Hence,asinusoidalexcitationtoaninertialsystemselectsthecorrespondingfrequencycomponentfromtransferfunctionofthesystem.Thoughsinusoidalexcitationiswidelyusedtoanalyzesyntheticsystems,itisnotcommonlyfoundinphysicalsystems.Randomexcitation:Randomexcitationcanbeinter-pretedasimpulseexcitationofarbitraryamplitudeateveryinstantoftime.Sinceimpulseexcitationsarepresentoveralltheinstantsoftime,itisdifÞculttoobservethemfromtheoutputofthesystem.Randomexcitationdoesnotpossessimpulse-likenatureeitherinthetime-domainorinthefrequency-domain,andhence,theimpulsescannotbeperceived.Sequenceofimpulsesasexcitation:Inthiscase,thesig-nalsaregeneratedbyapassiveinertialsystemwithaÞxedsequenceof(periodicand/oraperiodic)impulsesasexcitation.Thetimeinstantsofimpulsesmaynotbeobservedfromtheoutputofthesystem,buttheycanbeperceived.Ifthesequenceofimpulsesisperiodicinthetime-domain,thenitcorrespondstoaperiodicsequence Fig.1.Inertialsystemexcitedwithasequenceofimpulses.ofimpulsesinthefrequency-domainalso,andcanalsobeperceived.Consideraphysicalsystemexcitedbyasequenceofimpulsesofvaryingstrengths,asshowninFig.1.OneofthechallengesintheÞeldofsignalprocessingistodetectthetimeinstants oftheimpulsesandtheircorrespondingstrengths fromtheoutputsignal.Inanaturalscenariolikespeechproduction,thecharacteristicsofthesystemvarywithtimeandareunknown.Hence,thesignalprocessingproblemcanbeviewedasablinddeconvolution,whereneitherthesystemresponsenortheexci-tationsourceareknown.Inthispaper,weattempttodetectthetimeinstantsofexcitation(epochs)ofthevocal-tractsystem.Consideraunitimpulseinthetimedomain.Ithasallthefrequenciesequallywellrepresentedinthefrequencydomain.Whenaninertialsystemisexcitedbyanimpulse-likeexcitation,theeffectoftheexcitationspreadsuniformlyinthefrequencydomainandismodulatedbythetime-varyingtransferfunctionofthesystem.Theinformationaboutthetimeinstantsofoc-currenceoftheexcitationimpulsesreßectsasdiscontinuitiesinthetimedomain.ItmaybedifÞculttoobservethesedisconti-nuitiesdirectlyfromthesignalbecauseofthetime-varyingre-sponseofthesystem.TheeffectofthediscontinuitiescanbehighlightedbyÞlteringtheoutputsignalthroughanarrowbandÞltercenteredaroundafrequency.Theoutputofthenarrow-bandÞlterpredominantlycontainsasinglefrequencycompo-nent,andasaresult,thediscontinuitiesduetotheexcitationimpulseswillgetmanifestedasadeviationfromthecenterfre-quency.ThetimeinstantsofthediscontinuitiescanbederivedbycomputingtheinstantaneousfrequencyoftheÞlteredoutput[24].Atutorialreviewontheinstantaneousfrequencyanditsinterpretationisgivenin[25].Ithasbeenpreviouslyobservedthatisolatednarrowspikesintheinstantaneousfrequencyofthebandpass-Þlteredoutput[26,ch.11]areattributedtoeitherthevalleysintheamplitudeenvelopeortheonsetofanewpitchpulse,butnopreviousworkexploredthefeasibilityofthistypeofobservationforepochextraction.A.ComputationofInstantaneousFrequencyTheinstantaneousfrequencyofarealsignal isdeÞnedasthetimederivativeoftheunwrappedphaseofthecomplexanalyticsignalderivedfrom [24].Thecomplexanalyticsignalcorrespondingtoarealsignal isgivenby (1)where istheHilberttransformoftherealsignal isgivenby (2) MURTYANDYEGNANARAYANA:EPOCHEXTRACTIONFROMSPEECHSIGNALSwhereIFTdenotestheinverseFouriertransform,and givenby Theanalyticsignalthusderivedcontainsonlypositivefre-quencycomponents.Theanalyticsignal canberewritten (4)where iscalledtheamplitudeenvelope,and iscalledtheinstantaneousphase.Directcomputationofthe from(6)suffersfromtheproblemofphasewrapping, isconstrainedtoaninterval or .Hence,theinstantaneousfrequencycannotbecomputedbyexplicitdif-ferentiationofphase withoutÞrstperformingthecom-plextaskofunwrappingthephaseintime.Theinstantaneousfrequencycanbecomputeddirectlyfromthesignal,withoutgoingthroughtheprocessofphaseunwrapping,byexploitingtheFouriertransformrelations.Takinglogarithmonbothsidesof(4),anddifferentiatingwithrespecttotime ,wehave wherethesuperscript denotesthederivativeoperator,and istheinstantaneousfrequency.Thatis (8)where denotestheimaginarypart. canbecomputedbyusingtheFouriertransformrelations.Theanalyticsignal canbesynthesizedfromitsfrequencydomainrepresen-tationthroughtheinverseFouriertransform (9)where istheFouriertransformoftheanalyticsignal ,andiszerofornegativefrequencies.Differentiatingbothsidesof(9)withrespecttotime ,wehave Theinstantaneousfrequency canbeobtainedfrom(7)and(10)as (11)where denotesrealpart.Computationoftheinstantaneousfrequencygivenin(11)isimplementedinthediscretedomainasfollows: Here,IDFTdenotestheinversediscreteFouriertransform,and isthetotalnumberofsamplesinthesignal.Theinstantaneousfrequencymaybeinterpretedasthefre-quencyofasinusoidwhichlocallyÞtsthesignalunderanalysis.However,ithasaphysicalinterpretationonlyformonocom-ponentsignals,wherethereisonlyonefrequencyoranarrowrangeoffrequenciesvaryingasafunctionoftime.Inthiscase,theinstantaneousfrequencycanbeinterpretedasdeviationoffrequencyofthesignalfromthemonotoneateveryinstantoftime.Thenotionofasingle-valuedinstantaneousfrequencybe-comesmeaninglessformulticomponent(multiplefrequencysi-nusoids)signals.Themulticomponentsignalhastobedispersedintoitscomponentsforfurtheranalysis.Inthispaper,weproposetousearesonatortoÞlteroutfromasignalamonocomponentcenteredaroundasinglefrequencyforfurtheranalysis.Aresonatorisasecond-orderinÞnite-im-pulseresponse(IIR)Þlterwithacomplexconjugatepairofpolesinthe -plane[27].Aresonatorwithnarrowbandwidth(cor-respondingtoaradius )waschosentorealizethenarrowbandÞlter.Anidealresonator wasnotusedinordertoavoidsaturationoftheÞlteroutput.WhenamulticomponentsignalisÞlteredthrougharesonatorcenteredaroundafrequency ,theoutputsignalpredom-inantlycontainsthe frequencycomponent.Anydeviation infrequencyoftheÞlteredoutputcanbeattributedtotheimpulse-likecharacteristicspresentinthemulticomponentsignal.Ingeneral,theanalyticsignalcorrespondingtotheÞl-teredoutputcanbeexpressedas Hence,theinstantaneousfrequencyoftheÞlteredoutput(pre-dominantlymonocomponent)isgivenby Fig.2(a)showsamulticomponentsignalintheformofanape-riodicsequenceofimpulseswitharbitrarystrengths.ThesignalÞlteredbya500-Hzresonator,andtheinstantaneousfrequencyoftheÞlteredoutputareshowninFig.2(b)and(c),respec-tively.Attheinstantsofimpulselocations,theinstantaneousfrequencydeviatessigniÞcantlyfromthenormalizedcenterfre-quency ,where isthefrequencyoftheresonator, isthesamplingfrequency.Foraresonatorfrequency Hz,andsamplingfrequency ,theinstanta-neousfrequency(around )showssharppeaksatthelocationsoftheimpulses.TheillustrationinFig.2showsthatthediscontinuityinformationcanbederivedfromtheÞlteredoutputevenifthesequenceofimpulsesarenotregularlyspaced, Authorized licensed use limited to: INTERNATIONAL INSTITUTE OF INFORMATION TECHNOLOGY. Downloaded on January 30, 2009 at 00:56 from IEEE Xplore. Restrictions apply. MURTYANDYEGNANARAYANA:EPOCHEXTRACTIONFROMSPEECHSIGNALS Fig.5.A100-mssegmentof(a)speechwaveform,(b)outputoftheresonatorat500Hz,(c)instantaneousfrequencyoftheÞlteredoutput,and(d)differencedEGGsignal.bythedifferencedEGGsignal,illustratingthepotentialoftheproposedmethod.Inthecaseofspeech,instantaneousfrequencyoftheÞlteredoutputalsocontainsthetime-varyingfrequencychangesassoci-atedwiththevocal-tracttransferfunction,whichisundesirable.Asaresult,thoughthepeaksintheinstantaneousfrequencyoftheÞlteredoutputindicatetheepochlocationsaccuratelyforthesegmentshowninFig.5,itmaynotbeusefultoextracttheepochlocationsunambiguouslyforanychosencenterfrequency .Thus,themethodofepochextractionusingtheinstan-taneousfrequencyoftheÞlteredoutputdependscriticallyonthechoiceofcenterfrequencyoftheÞlter.Asinglecenterfre-quencymaynotbesuitableforextractingtheepochlocationsofanarbitrarysegmentofspeech.Thecenterfrequencyhastobechosenbasedonthecharacteristicsofthespeechsegmentunderanalysis.Thechoiceofthecenterfrequencyalsodependsonthedistributionofenergyofthespeechsegmentinthefrequencydomain.ToillustratethesigniÞcanceofchoiceofthecenterfrequencyoftheÞlter,theinstantaneousfrequencycomputedaroundfourdifferentcenterfrequenciesareshowninFig.6.Thespectrogram,thespeechsignalandthedifferencedEGGsignalarealsogivenforreference.ThespectrograminFig.6(a)showsabandofenergyaround500Hz.Theinstantaneousfrequencycomputedaround500Hz[Fig.6(d)]indicatesunambiguouspeaks/valleysthatareincloseagreementwiththeactualepochsshownbythedifferencedEGGsignal[Fig.6(c)].Intheinstanta-neousfrequenciescomputedaround1000and2000Hz,showninFig.6(e)and(f),respectively,theepochlocationscannotbeidentiÞedeasily.Thisisbecausetheenergyofthesignalinthosefrequencybandsisverylow.Sincethespectrogramshowslargeenergyinthebandaround2500Hz,theinstantaneousfrequencycomputedaround2500Hzshowssharppeaks/valleysaroundtheepochlocations.However,theinstantaneousfrequencyplotinFig.6(g)showslessambiguouspeaks/valleysinthetimein-terval570Ð620ms,thanthoseinthetimeinterval520Ð570ms.Thisisbecausetheintensityofthe2500-Hzfrequencybandinthetimeinterval570Ð620msisgreaterthantheintensityofthebandinthetimeinterval520Ð570ms. Fig.6.Instantaneousfrequencyofaspeechsegmentcomputedaroundfourdifferentcenterfrequencies.(a)Spectrogramofthespeechsegment.(b)Speechwaveform.(c)DifferencedEGGsignal.Instantaneousfrequencyplotscom-putedaround(d)500Hz,(e)1000Hz,(f)2000Hz,and(g)2500Hz.Noticethattheinstantaneousfrequenciescomputedaround1000and2000Hzalsocontainallthepeaks/valleyscorre-spondingtotheepochlocations,buttheycannotbelocatedeasilyduetoßuctuationsintheneighborhood.Thisisbecausetheinstantaneousfrequencycapturesnotonlythediscontinu-itiesduetotheexcitationimpulses,butalsotheßuctuationsduetothetime-varyingvocal-tractsystem.Hence,itisdifÞculttoextracttheinstantsofexcitationfromtheinstantaneousfrequencycomputedaroundanarbitrarycenterfrequency.Thecenterfrequencyhastobechoseninsuchawaythatthediscontinuitiesduetotheexcitationimpulsesdominateovertheßuctuationsduetothetime-varyingvocal-tractsystem.IV.EXTRACTIONROMSINGA0-HzRESONATORThediscontinuityduetoimpulseexcitationisreßectedacrossallthefrequenciesincludingthezerofrequency.Thatis,eventheoutputoftheresonatoratzerofrequencyshouldhavethein-formationofthediscontinuitiesduetoimpulse-likeexcitation.TheadvantageofchoosingthezerofrequencyresonatorÞlteristhatthecharacteristicsofthetime-varyingvocal-tractsystemwillnotaffectthecharacteristicsofthediscontinuitiesintheres-onatorÞlteroutput.Thisisbecausethevocal-tractsystemhasresonancesatmuchhigherfrequenciesthanatzerofrequency. Authorized licensed use limited to: INTERNATIONAL INSTITUTE OF INFORMATION TECHNOLOGY. Downloaded on January 30, 2009 at 00:56 from IEEE Xplore. Restrictions apply. Authorized licensed use limited to: INTERNATIONAL INSTITUTE OF INFORMATION TECHNOLOGY. Downloaded on January 30, 2009 at 00:56 from IEEE Xplore. Restrictions apply. 1610IEEETRANSACTIONSONAUDIO,SPEECH,ANDLANGUAGEPROCESSING,VOL.16,NO.8,NOVEMBER2008 Fig.12.Illustrationofgroup-delaybasedmethodforepochextraction[18].(a)Speechsignal,(b)LPresidual,(c)phase-slopefunction,and(d)differencedEGGsignal.Thepulsesin(d)indicatethedetectedepochlocations.fromthespeechsignalalone.Therearethreecomponentsinthealgorithm.TheÞrstcomponentgeneratescandidateepochsusingzerocrossingsofthephase-slopefunction.Theenergyweightedgroup-delaywasusedasameasuretoderivethephase-slopefunction.Thesecondcomponentemploysanovelphase-slopeprojectiontechniquetorecovercandi-datesforwhichthephase-slopefunctiondoesnotincludeazero-crossing.Thesetwocomponentsdetectalmostallthetrueepochs,buttheyalsogeneratealargenumberoffalsealarms.Thethirdcomponentofthealgorithmusesdynamicprogrammingtoidentifythetrueepochsfromthesetofhypoth-esizedcandidatesbyminimizingacostfunction.Forevaluatingthistechnique,theMATLABimplementationoftheDYPSAavailablein[29]wasused.TheCMU-Arcticdatabase[30],[31]wasemployedtoeval-uatetheproposedmethodofepochdetectionandtocomparetheresultswiththeexistingmethods.TheArcticdatabasecon-sistsof1132phoneticallybalancedEnglishsentencesspokenbytwomaleandonefemaletalkers.Thedurationofeachutter-anceisapproximately3s,whichmakesthedurationoftheentiredatabasetobearound2h40min.Thedatabasewascollectedinasoundproofbooth,anddigitizedatasamplingfrequencyof32kHz.Inadditiontothespeechsignals,theArcticdata-basecontainsthesimultaneousrecordingsofEGGsignalscol-lectedusingalaryngograph.ThespeechandEGGsignalsweretime-alignedtocompensateforthelarynx-to-microphonedelay,determinedtobeapproximately0.7ms.ReferencelocationsoftheepochswereextractedfromthevoicedsegmentsoftheEGGsignalsbyÞndingpeaksinthedifferencedEGGsignal.Theper-formanceofthealgorithmswasevaluatedonlyinthevoicedseg-ments(detectedfromEGGsignal)betweenthereferenceepochlocationsandtheestimatedepochlocations.Thedatabasecon-tainsatotalof792249epochsinthevoicedregions.Theperformanceoftheepochdetectionmethodswasevalu-atedusingthemeasuresdeÞnedin[23].Fig.13showsthechar-acterizationofepochestimatesshowingeachofthepossiblede-cisionsfromtheepochdetectionalgorithms.Thefollowing Fig.13.Characterizationofepochestimatesshowingthreelarynxcycleswithexamplesofeachpossibleoutcomefromepochextraction[23].IdentiÞcationaccuracyismeasuredasavarianceofTABLEIOMPARISONOFETHODSONATABASE.IDRÑIDENTIFICATIONATE,MRÑMATEFARÑFATE,IDAÑIDENTIFICATIONCCURACY measuresweredeÞnedtoevaluatetheperformanceofepochdetectionalgorithms.Larynxcycle:Therangeofsamples ,givenanepochreferenceat withprecedingandsucceedingepochrefer-encesatsamples and ,respectively.IdentiÞcationrate(IDR):Thepercentageoflarynxcy-clesforwhichexactlyoneepochisdetected.Missrate(MR):Thepercentageoflarynxcyclesforwhichnoepochisdetected.Falsealarmrate(FAR):Thepercentageoflarynxcyclesforwhichmorethanoneepochisdetected.IdentiÞcationerror :Thetimingerrorbetweentheref-erenceepochlocationandthedetectedepochlocationinlarynxcyclesforwhichexactlyoneepochwasdetected.IdentiÞcationaccuracy (IDA):ThestandarddeviationoftheidentiÞcationerror .Smallvaluesof highaccuracyofidentiÞcation.TableIshowstheperformanceresultsonArcticdatabaseforidentiÞcationrate,missrate,falsealarmrate,andidentiÞca-tionaccuracyforthethreemethodsHE-based,GD-based,andDYPSAalgorithm,aswellasfortheproposedmethod.Fig.14showsthehistogramsofthetimingerrors indetectingtheepochlocations,averagedovertheentireArcticdatabase.FromTableI,itcanbeconcludedthattheDYPSAalgorithmper-formedbestamongthethreeexistingtechniques,withaniden-tiÞcationrateof96.66%.Theproposedmethodofepochde-tectiongivesabetteridentiÞcationrateaswellasidentiÞcationaccuracy,comparedtotheresultsfromtheDYPSAalgorithm. MURTYANDYEGNANARAYANA:EPOCHEXTRACTIONFROMSPEECHSIGNALS Fig.14.Histogramoftheepochtimingerrorsfor(a)HE-basedmethod,(b)GD-basedmethod,(c)DYPSAalgorithm,and(d)proposedmethod. Fig.15.HistogramoftheepochtimingerrorsfordegradationbywhitenoiseatanSNRof10dB.(a)HE-basedmethod,(b)GD-basedmethod,(c)DYPSAalgorithm,and(d)proposedmethod.VI.EFFECTOFOISEONERFORMANCEOFROPOSEDETHODOFXTRACTIONInthissection,westudytheeffectof(moderatelevelsof)noiseontheaccuracyoftheepochdetectionmethods.Theex-istingmethodsandtheproposedmethodareevaluatedonanartiÞciallygeneratednoisyspeechdatabase.Severalnoiseen-vironmentsatvaryingsignal-to-noiseratio(SNR)weresimu-latedtoevaluatetheepochdetectionmethods.ThenoiseusedwastakenforNOISEX-92database[32].Thedatabasecon-sistsofwhite,babble,high-frequency(HF)channel,andvehiclenoise.ThenoisefromtheNOISEX-92databasewasaddedtotheArcticdatabasetoformnoisyspeechdataatdifferentlevelsofdegradation.Theutterancesareappendedwithsilencesuchthatthetotalamountofsilenceineachutteranceisconstrainedtobeabout60%ofdata,includingthepausesintheutterances.IncludingdifferentnoiseenvironmentsandSNRs,thedatabaseconsistsof33hofnoisyspeechdata.TableIIshowsthecomparativeresultsofepochdetectionmethodsfordifferenttypesofdegradationsatvaryingSNRs.Fig.15showsthedistributionofthetimingerrors indetectingtheepochlocations,forwhitenoiseenvironmentatanofSNRof10dB.Theproposedmethodconsistentlyperformsbetterthantheexistingtechniquesevenunderdegradation.Theimprovedperformanceoftheproposedmethodmaybeattributedtothefollowingreasons.1)Thereisnoblockprocessinginvolvedinthismethod.Hence,therearenoeffectsofthesizeandtheshapeofthewindow.TheentirespeechsignalisprocessedatoncetoobtaintheÞlteredsignal.2)Theproposedmethodisnotdepen-dentontheenergyofthesignal.Thismethoddetectstheepochlocationseveninweaklyvoicedregionslikevoice-bar.3)Thereisonlyoneparameterinvolvedintheproposedmethod,i.e.,thelengthofthewindowforremovingthetrendfromtheoutputof0-Hzresonator.4)Therearenocriticalthresholdsorcostsin-volvedinidentifyingtheepochlocations.VII.SUMMARYANDInthispaper,weproposedamethodforepochextractionthatdoesnotdependonthecharacteristicsofthevocal-tractsystem.Themethodexploitstheimpulse-likeexcitationofthevocal-tractsystem.Themethodusestheoutputofspeechfromazerofrequencyresonator.ThepositivezerocrossingsoftheÞlteredsignalcorrespondtoepochs.TheidentiÞcationrateandidenti-ÞcationaccuracyareevaluatedusingtheCMU-Arcticdatabase,wherethespeechsignalandthecorrespondingEGGsignalsareavailable.TheepochinformationderivedfromtheEGGsig-nalsisusedasareference.TheperformanceoftheproposedmethodiscomparedwiththeresultsfromtheDYPSAandtwoothermethods.TheproposedmethodgivessigniÞcantlybetterresultsintermsofidentiÞcationrateandidentiÞcationaccu-racy.Itisalsointerestingtonotethattheproposedmethodisro-bustagainstdegradationssuchaswhitenoise,babble,high-fre-quencychannel,andvehiclenoise.Therearemanynovelfeaturesintheproposedmethodofepochextraction.Themethoddoesnotuseanyblockprocessingasmostsignalprocessingmethodsdo.Theperformanceofthemethoddoesnotdependontheenergyofthesegmentofspeechsignal,andhence,themethodworksequallywellforalltypesofspeechsoundunits.Inaddition,therearenoparameterstocon-trol,andnoarbitrarythresholdingintheidentiÞcationofepochs. 1612IEEETRANSACTIONSONAUDIO,SPEECH,ANDLANGUAGEPROCESSING,VOL.16,NO.8,NOVEMBER2008TABLEIIOMPARISONFORETHODSFORSANDNVIRONMENTSDENTIFICATIONATE,MRÑMATE,FARÑFATE,IDAÑIDENTIFICATIONCCURACY Themethodperformswellforspeechcollectedwithaclose-speakingmicrophone,evenwiththeadditionofdegradations.However,themethodisnotlikelytoworkwellwhenthedegra-dationsproduceadditionalimpulse-likesequencesinthecol-lectedspeechdataasinthecaseofreverberation.Themethodisalsonotlikelytoworkwellwhenthereisinterferenceofspeechfromotherspeakers.Ourfutureeffortswillbeinthedirectionofdevelopingmethodsforextractingepochsfromspeechwithdegradationsinvolvingsuperposedimpulse-likecharacteristics.Sincetheproposedmethodprovidesaccuratelocationsofepochs,theresultsareusefultodevelopmethodsforpitchextraction,andalsoforvoiceactivitydetection.Also,sincetheÞlteredsignalgivesanindicationofglottalactivity,themethodmaybeusedforanalysisofphonationcharacteristics[33]innormalandpathologicalvoices.ThemethodmayalsobeausefulÞrststepinaccurateanalysisofvocal-tractcharacteristicsbyfocusingtheattentionintheregionaroundtheepochs.Accu-rateanalysisofexcitationsourceandtime-varyingvocal-tractsystemsmayleadtoabetteracousticÐphoneticanalysisofspeechsoundsinmanylanguages,anditalsomayprovideausefulsupplementtotheexistingspectral-basedmethodsofspeechanalysis. Authorized licensed use limited to: INTERNATIONAL INSTITUTE OF INFORMATION TECHNOLOGY. Downloaded on January 30, 2009 at 00:56 from IEEE Xplore. Restrictions apply.