Fig.1.Laryngealparametersinthearticulatorydescriptionofphonationtypes[ - PDF document

Fig.1.Laryngealparametersinthearticulatorydescriptionofphonationtypes[16].Section4,methodsusedforanalysisareexplained.InSection5,resultsoftheanalysisarediscussed.Finally,summaryandconclusionsarepresentedinSection6.II.PRODUCTIONMECHANISMOFBREATHYVOICESHumanbeingscanproducespeechsoundswithnotonlyreg-ularvoicingvibrationsatarangeofdifferentpitchfrequencies,butalsowithavarietyofvoicesourcecharacteristicsreectingdifferentvoicequalities.Thevoiceiscontrolledbydifferenttypesofmusculartensions,namely,adductivetension,medialcompressionandlongitudinaltension[15].Adductivetensioniscontrolledbyinterarytenoidmusclesanddrawsthearytenoidstogether.Medialcompressioniscontrolledbythelateralcricoarytenoidmusclesandkeepstheligamentalglottisclosed.Similarly,longitudinaltensionismediatedprimarilybymusclesofthevocalfoldsandthecricothyroidmuscles.Thecontractionofthecricoarytenoidmusclescanalsoincreasethelongitudinaltensionbytiltingthearytenoidcartilagesbackwards.Figure1showsthelaryngealparametersinthearticulatorydescriptionofphonationtypes[16].Breathyvoicecanbeproducedbymaintaininganopenglottisforthemajorityorentiretyofthevibrationcycle,oritcanbecausedbyvocalfoldswhichclosemoreslowlythanformodalphonation.Breathinessischaracterizedbylowadductivetensionandmoderatetohighmedialcompression.Atriangularopeningbetweenthearytenoidcartilages,whichisaconsequenceofthelowadductivetensioncombinedwiththecongurationofvoicing,producesabreathyvoicedphonation.Itischaracterizedbyvocalfoldshavinglittlelongitudinaltension.Thisresultsinsometurbulentairowthroughtheglottis.Thuswehavetheauditoryimpressionof“voicemixedinwithbreath”[5].Breathyvoiceandwhisperaredifferent.Theydifferinthemannerofproduction.Inbreathyvoicethevocalmuscletensionislowandthereisvoicing,whereasinwhisperthevocalmuscletensionishighandthereisnovoicinginvolved.III.DATACOLLECTIONInitialdatafortheanalysiswascollectedat48kHzsamplingfrequencyinaquietroom,andwasrecordedbyanexpertphonetician.Datawasalsorecordedfrom10nativeGujaratispeakers(8maleand2female).TheparticipantsincludethosewhohavestayedmostofthetimeinGujarat,andthosewhohaven'tbeentherebutarenativesofthatlanguage.Syllablescontainingthebreathyvoicephonationinthevowelpartwererecordedbyanexpertphonetician.Somewordscontainingthebreathyphonationandtheircorrespond-ingcontrastingmodalphonatedwordswererecordedbythenativeGujaratispeakers.Toensureuniformprosodiceffect,andtohelpthespeakerstospeaknaturally,meaningfuldeclar-ativecarrierwordsandsentenceswereused.IV.PARAMETERS/FEATURESFORBREATHYVOICEFollowingarethetechniquesusedtocomputetheparam-etersfortheanalysisofthebreathyvoicedsignals.Thesetechniquesarerobustbecausetheyattempttocapturetheacousticpropertiesoftheactualspeechproductionmechanism.A.Zero-frequencylteringtechniqueAmethodisproposedforextractionoftheinstantaneousF0,epochextractionandstrengthofimpulse-likeexcitationatepochs[17],[18].Themethodusesthezero-frequencylteredsignalderivedfromspeechtoobtaintheepochs(instantsofsignicantexcitationofthevocaltractsystem)andthestrengthofimpulseattheepochs.Themethodinvolvespassingthedifferencedspeechsignalthroughacascadeoftwoidealdigitalresonators,eachlocatedat0Hz.Thetrendintheoutputisremovedbysubtractingthelocalmeanateachsample,computedoverawindowlengthintherangeofabout1to2pitchperiods.Thenegativetopositivezerocrossinginstantsintheresultingzerofrequencyltered(ZFF)outputarecalledepochs.TheslopesoftheZFFsignalatepochsgivetherelativestrengthsoftheimpulse-likeexcitation(SoE)aroundepochs.Thereciprocaloftheintervalbetweensuccessiveepochsgivestheinstantaneousfundamentalfrequency(F0).ItisobservedthattheF0ofaspeakerisloweredduringbreathyphonationwhencomparedtoF0duringmodalvoiceasshowninFigures2and3.Thedecreaseintheoverallfundamentalfrequencyvaluescanbeattributedtothefactthatduringthebreathyvoice,thereisagapforairtoowthroughthevocalfoldsandthisresultsintheslowervibrationsofthevocalfolds.BreathyphonationstartswithlowerF0andincreasessteeplyinashortduration.Theriseisashighas20%inlessthanaperiodof10milliseconds.WealsoobservefromFigure3thatthereisasuddenriseintheSoEfromthestopconsonanttothevowelinthemodalvoicesignalasanticipated,whereasweobserveinFigure2thatthetransitionintheSoEforbreathyvoiceisgradual.Thisisbecausethereisnoabruptnessintheglottalclosuremechanismofabreathyphonation. Fig.2.Breathyvoicedsyllable(/bi:/)'s(a)Waveform,(b)Instantaneousfundamentalfrequencyand(c)StrengthofExcitation. Fig.3.Modalvoicedsyllable(/bi/)'s(a)Waveform,(b)Instantaneousfundamentalfrequencyand(c)StrengthofExcitation.B.Periodic-aperiodicenergycomputationBreathyvoicespeechconsistsofincreasedspectralnoise,particularlyathigherfrequencies.Thisisduetopersistentleakageofairthroughtheglottisduringbreathyphonation.Theratiooftheperiodicandaperiodicenergies(PAP)canbeusedasameasuretoreectthisproperty.ThisapproachtocalculatePAPinvolvesiterativedecom-positionofspeechintoperiodicandaperiodiccomponentsasproposedin[19].Themethodissummarizedinthefollowingsteps:(a)Performlinearprediction(LP)analysistocomputetheLP-residual(b)DividetheLPresidualintoframesofsize32mswithaframeshiftof4ms.Checkforvoicedandunvoicedframes.(c)Computecepstrumusing512pointFFTandHammingwindow.Identifythepeakincepstrumrelatingtohar-monicsinspectrumbyusingpitchinformationobtainedbyZFFmethod(Section3.1). Fig.4.PAPformodalvoice(/bi/):(a)Waveform,(b)Periodicenergy,(c)Aperiodicenergyand(d)PAPratio. Fig.5.PAPforbreathyvoice(/bi:/):(a)Waveform,(b)Periodicenergy,(c)Aperiodicenergyand(d)PAPratio.(d)Computetheharmoniclogspectrumbymakingallthecoefcientsincepstrum,exceptthe9samplesaroundthepeakcorrespondingtothepitchperiodtozero,andtakeIDFT.(e)ComputethespectrumoftheLPresidualframe.Samplesfromthespectrumarenowdividedintoperiodicandaperiodicparts.(f)Aniterativealgorithmisusedtocomputetheaperiodiccomponentoftheresidual.Periodiccomponentisob-tainedbysubtractingtheaperiodiccomponentfromtheresidualofthespeechsignal.(g)Synthesizeperiodicandaperiodiccomponentsofthespeechsignalbyexcitingtheallpolelter(LPsynthesis)withtheperiodicandaperiodiccomponentsoftheresidualasexcitation,respectively.Theratioofenergyoftheperiodicandaperiodiccompo-nents(Ep Eap)computedovereachoftheframesinthevoiced breathyandmodalvoiceislesserthanthecontrastobservedforanaturallynon-breathyvoice.TABLEIIITableshowingthemeanvaluesoftheparametersforbreathyandmodalsounds. Parameter Breathy Modal F0(Hz) 114 123 SoE 0.0118 0.0101 PAP 9.93 19.00 Loudness 0.52 0.70 A1-A2(dB) 2.61 2.21 A1-A3(dB) 3.25 2.84 BothconventionalfeaturessuchasspectraltiltandnewfeatureslikePAP;loudness;SoEareusedtodescribetheacousticcharacteristicsofbreathyvoicequality.Thesefeaturescanbeusedtospotbreathinessinaspeechsignal.Weobservethatbreathyvoiceisperceivedtobelessloudthanthemodalvoiceanditismeasuredbyaloudnessmeasure.Duetohigheramountofaperiodicityattachedwithbreathyphonation,weseethatthePAPratioislessforbreathyvoicequality.TheaverageF0islessandthestrengthofexcitationismoreforbreathyvoice.ThespectraltiltishigherforbreathyvoiceasconrmedbythemeasuresofA1-A2andA1-A3.VI.SUMMARYANDCONCLUSIONSInthispaperwehaveemphasizedtheneedtolookintofeaturesbasedonexcitationcharacteristicstoderiveacousticcuesforbreathyvoices.Featuresbasedoninstantaneousfundamentalfrequencycontour,thestrengthofimpulse-likeexcitationatepochs,theratiooftheperiodicandaperiodiccomponentsofspeechandloudnessmeasurederivedfromtheHEoftheLPresidualwereproposedtosupplementthespectralfeaturesforcharacterizingbreathyvoices.Thederivedloudnessmeasurewasusedtoclassifybreathyvowels.Thestudiesreportedinthispaperhelptoderivefeaturesofbreathyvoiceforspottingsuchsegmentsincontinuousspeech.REFERENCES[1]PeterLadefogedandIanMaddieson,“TheSoundsoftheWorld'sLan-guages”.Oxford:Blackwell.1996.[2]J.HillenbrandandR.A.Houde,“AcousticCorrelatesofBreathyVocalQuality:DysphonicVoicesandContinuousSpeech,”JSpeechHearRes,vol.39,pp.311-321,Apr.1996.[3]MeysamAsgariandIzhakShafran,“ExtractingcuesfromSpeechforpredictingseverityofParkinson'sdisease”,inInternationalWorkshoponMachinesLearningforSignalProcessing,Kittila,Finland,2010.[4]BarbaraBlankenship,“Thetimingofnonmodalphonationinvowels”,JournalofPhonetics,vol.30,no.2,pp.163-191,2002.[5]MatthewGordonandPeterLadefoged,“PhonationTypes:across-linguisticoverview”,JournalofPhonetics,29,383-406,2001.[6]RatreeWaylandandAllardJongman,“Acousticcorrelatesofbreathyandclearvowels:thecaseofKhmer”,JournalofPhonetics,31,181-201,2003.[7]HansonH.,“Glottalcharacteristicsoffemalespeakers:Acousticcorre-lates”,J.Acoustic.Soc.Amer.,Vol.101:466-481.,1997[8]Alku,P.,Vilkman,E.,“Amplitudedomainquotientforcharacterizationoftheglottalvolumevelocitywaveformestimatedbyinverseltering”,SpeechCommunication,Vol.18(2),131-138,1996[9]M.Fr¨ohlich,D.Michaelis,andH.W.Strube,“Acoustic“breathinessmeasures”inthedescriptionofpathologicalvoices,”inProc.ICASSP,vol.2,pp.937-940,Seattle,WA,May1998.[10]Michaelis,D.,Gramss,T.,Strube,H.W.,“Glottal-to-noiseexcitationratio-anewmeasurefordescribingpathologicalvoices”,Acustica,Vol.83,700-706,1997[11]IshiC.T.,IshiguroH.,HagitaN.,“Analysisoftherolesandthedynamicsofbreathyandwhisperyvoicequalitiesindialoguespeech”,EURASIPJournalonAudio,Speech,andMusicProcessing2010,ID528193,1-12,Jan.2010.[12]IshiC.T.,IshiguroH.,HagitaN.,“ImprovedAcousticCharacterizationofBreathyandWhisperyVoices”InProc.ofInterspeech2011,pp.1237-1240.[13]Ji-YeounLee,SangbaeJeong,MinsooHahnandHong-ShikChoi,“Automaticvoicequalitymeasurementbasedonefcientcombinationofmultiplefeatures”,inICBBE,pp.1272-1275,May2008.[14]EduardoCastillo-Guerra,AdelRuz,“Automaticmodelingofacousticperceptionofbreathinessinpathologicalvoices”,IEEETransactionBiomedicalEngineeringvol.56(4),pp.932-940,2009[15]C.Gobl,“Apreliminarystudyofacousticvoicequalitycorrelates”,STL-QPSR,vol.4,pp.9-21,1989.[16]Laver,J.,“ThePhoneticDescriptionofVoiceQuality”,CambridgeUniversityPress,Cambridge.1980.[17]K.SriRamaMurthyandB.Yegnanarayana,“EpochExtractionfromSpeechSignals,”IEEETrans.Audio,SpeechLang.Process.,vol.16,no.8,pp.1602-1613,Nov.2008.[18]K.SriRamaMurtyandB.YegnanarayanaandAnandJosephM.,“CharacterizationofGlottalActivityfromSpeechSignals,”IEEEsignalprocessingletters,vol.16,no.6,June2009.[19]B.Yegnanarayana,C.R.d'Alessandro,andV.Darsinos,“Aniterativealgorithmfordecompositionofspeechsignalsintoperiodicandaperiodiccomponents”IEEETrans.SpeechAudioProcessing,vol.6,pp.1-11,Feb.1998.[20]AnandJosephM.,GuruprasadS.andYegnanarayanaB.,“ExtractingFormantsfromShortSegmentsofSpeechusingGroupDelayFunc-tions”,Proc.Int.Conf.SpokenLanguageProcessing(INTERSPEECH)2006,pp.1009-1012,PittsburghPA,USA,17-21Sept.2006.[21]G.SheshadriandB.Yegnanarayana,“Perceivedloudnessofspeechbasedonthecharacteristicsofexcitationsource”,JournalofAcousticalSocietyofAmerica,Vol.126,No.4,pp.2061-2071,Oct2009.