共振器googlevoice

Chapter1IntroductiontoSpeechSignalProcessing语音信号处理概述1OutlineTheSpeechSignalSpeechSignalProcessingSpeechProduction/PerceptionModelandtheSpeechChainTheSpeechStackApplicationsofSpeechSignalProcessingHistoryofSpeechSignalProcessing2TheSpeechSignalSpeech(语音)isthevocalized(有声的)formofhumancommunicationThefundamentalpurposeofspeechishumancommunication;i.
e.
,thetransmissionofmessages(信息)betweenaspeakerandalistenerThefundamentalanalogformofthemessageisanacousticwaveform(声学波形)thatwecallthespeechsignal(语音信号)Speechsignalscanbe–convertedtoanelectricalwaveformbyamicrophone–manipulatedbyanalog/digitalsignalprocessing–convertedbacktoacousticformbyaloudspeaker/headphone3TheSpeechSignal4SoftwarePraat–http://www.
fon.
hum.
uva.
nl/praat/CoolEditPro(AdobeAudition)5SpeechSignalProcessingSpeechSignalProcessing(语音信号处理)–convertingonetypeofspeechsignalrepresentationtoanothersoastouncovervariousmathematicalorpracticalpropertiesofthespeechsignal(发掘语音特征)anddoappropriateprocessingtoaidinsolvingbothfundamentalanddeepproblemsofinterest(解决实际问题)Purposeofspeechsignalprocessing–Tounderstandspeechasameansofcommunication–Torepresentspeechfortransmissionandreproduction–Toanalyzespeechforautomaticrecognitionandextractionofinformation–Todiscoversomephysiologicalcharacteristicsofthetalker6SpeechSignalProcessingDigitalprocessingofspeechsignal(数字语音信号处理,DPSS)–obtainingdiscreterepresentationsofspeechsignal,whichpreservestheinformationcontentinthespeechsignal,alsoitisconvenientfortransmissionorstorage–theory,designandimplementationofnumericalprocedures(algorithms)forprocessingthediscreterepresentationinordertoachieveagoal(recognizingthesignal,modifyingthetimescaleofthesignal,removingbackgroundnoisefromthesignal,etc.
)7SpeechSignalProcessingAdvantagesofDPSS–reliability–flexibility–accuracy–real-timeimplementationsoninexpensiveDSPchips–abilitytointegratewithmultimediaanddata–encryptability/securityofthedataandthedatarepresentationsviasuitabletechniques8OutlineTheSpeechSignalSpeechSignalProcessingSpeechProduction/PerceptionModelandtheSpeechChainTheSpeechStackApplicationsofSpeechSignalProcessingHistoryofSpeechSignalProcessing9SpeechProductionModelMessageFormulation信息形成–desiretocommunicateanidea,awish,arequest,…expressthemessageasasequenceofwords10SpeechProductionModelLanguageCode语言编码–needtoconvertchosentextstringtoasequenceofsoundsinthelanguagethatcanbeunderstoodbyothers–needtogivesomeformofemphasis,prosody(tune,melody)tothespokensoundssoastoimpartnon-speechinformationsuchassenseofurgency,importance,psychologicalstateoftalker,environmentalfactors(noise,echo)11SpeechProductionModelNeuro-MuscularControls神经-肌肉控制–needtodirecttheneuro-muscularsystemtomovethearticulators(发音器官)(tongue,lips,teeth,jaws,velum(软腭))soastoproducethedesiredspokenmessageinthedesiredmanner12SpeechProductionModelVocalTract(声道)System–needtoshapethehumanvocaltractsystemandprovidetheappropriatesoundsourcestocreateanacousticwaveform(speech)thatisunderstandableintheenvironmentinwhichitisspoken13SpeechPerceptionModelTheacousticwaveformimpinges(冲击)ontheear(thebasilarmembrane(基底膜))andisspectrallyanalyzedbyanequivalentfilterbank(滤波器组)oftheearThesignalfromthebasilarmembraneisneurallytransducedandcodedintofeaturesthatcanbedecodedbythebrain14SpeechPerceptionModelThebraindecodesthefeaturestreamintosounds,wordsandsentencesThebraindeterminesthemeaningofthewordsviaamessageunderstandingmechanism15TheSpeechChain16Goal:FindoutifyourofficematehashadlunchText:"Didyoueatyet"Phonemes:"didyuityt"ArticulatorDynamics:dIjitjtInformationRateofSpeechText(discrete)–2^5symbols,10symbols/s->50bpsPhonemes&Prosody(discrete)–200bpsArticulatorymotions(continuous)–Relativelyslowmovementofarticulators~2000bpsAcousticwaveform(continuous)–64,000bps~705,600bps17TheSpeechStack18SpeechScience(语音科学)Linguistics(语言学):scienceoflanguage,includingsyntax,semantics,phonetics,phonology,etc.
Syntax(句法,语法):analysisanddescriptionofthegrammaticalstructureofabodyoftextualmaterialSemantics(语义学):analysisanddescriptionofthemeaningofabodyoftextualmaterialanditsrelationshiptoataskdescriptionofthelanguagePhonetics(语音学):studyofspeechsoundsandtheirproduction,transmission,andperception,andtheiranalysis,classification,andtranscription–Articulatory/Acoustic/AuditoryPhoneticsPhonology(音系学):systematicorganizationofsoundsinlanguages,systemsofphonemesinparticularlanguagesPhonemes(音位,音素):smallestsetofunitsconsideredtobethebasicsetofdistinctivesoundsofalanguages(20-60unitsformostlanguages)ApplicationsofSpeechSignalProcessingSpeechcoding(语音编码)Speechsynthesis(语音合成)Speechrecognitionandunderstanding(语音识别与理解)Otherspeechapplications20SpeechCodingTheprocessoftransformingaspeechsignalintoarepresentationforefficienttransmissionandstorageofspeech–narrowbandandbroadbandwiredtelephony–cellularcommunications–VoiceoverIP(VoIP)toutilizetheInternetasareal-timecommunicationsmedium–securevoiceforprivacyandencryptionfornationalsecurityapplications–extremelynarrowbandcommunicationschannels,e.
g.
,battlefieldapplicationsusingHFradio–storageofspeechfortelephoneansweringmachines,IVRsystems,prerecordedmessages21SpeechCoding22ApplicationsofSpeechSignalProcessing23SpeechSynthesisTheprocessofgeneratingaspeechsignalusingcomputationalmeansforeffectivehuman-machineinteractions–machinereadingoftextoremailmessages–telematicsfeedbackinautomobiles–talkingagentsforautomatictransactions–automaticagentincustomercarecallcenter–handhelddevicessuchasforeignlanguagephrasebooks,dictionaries,crosswordpuzzlehelpers–announcementmachinesthatprovideinformationsuchasstockquotes,airlines–schedules,weatherreports,etc.
24SpeechSynthesis25SpeechRecognitionandUnderstandingTheprocessofextractingusablelinguisticinformationfromaspeechsignalinsupportofhuman-machinecommunicationbyvoice–commandandcontrol(C&C)applications,e.
g.
,simplecommandsforspreadsheets,presentationgraphics,appliances–voicedictationtocreateletters,memos,andotherdocuments–naturallanguagevoicedialogueswithmachinestoenableHelpdesks,CallCenters–voicedialingforcellphonesandfromPDA'sandothersmalldevices–agentservicessuchascalendarentryandupdate,addresslistmodificationandentry,etc.
26PatternMatchingProblems27OtherSpeechApplicationsSpeakerVerification(话者确认)–forsecureaccesstopremises,information,virtualspacesSpeakerRecognition(话者识别)–forlegalandforensicpurposes—nationalsecurity;alsoforpersonalizedservicesSpeechEnhancement(语音增强)–foruseinnoisyenvironments,toeliminateecho,toalignvoiceswithvideosegments,tochangevoicequalities,tospeed-uporslow-downprerecordedspeech(e.
g.
,talkingbooks,rapidreviewofmaterial,carefulscrutinizingofspokenmaterial,etc)–potentiallytoimproveintelligibilityandnaturalnessofspeechLanguageTranslation(语言翻译)–toconvertspokenwordsinonelanguagetoanothertofacilitatenaturallanguagedialoguesbetweenpeoplespeakingdifferentlanguages,i.
e.
,tourists,businesspeople28HistoryofSpeechSignalProcessing29HistoryofSpeechSignalProcessingInventionoftelephone,Bell1876–"Watson,ifIcangetamechanismwhichwillmakeacurrentofelectricityvaryitsintensityastheairvariesindensitywhensoundispassingthroughit,Icantelegraphanysound,eventhesoundofspeech"30HistoryofSpeechSignalProcessingVOCODERandVODER,Dudley–VOCODER(VOiceenCODER)声码器amethodofreproducingspeechthroughelectronicmeanssource-filtermodeluseparallelband-passfiltertofilterspeechintotenspecificaudiospectrumbands,renderingitmoreeasilytransmittedovertelephonelines–VODER(VoiceOperationDEmonstratoR)aconsolefromwhichanoperatorcouldcreatephrasesofspeechcontrollingaVOCODERwithakeyboardandfootpedals(踏板)1939WorldFairinNYC31VODERVODERSoundSpectrograph(语谱仪),BellLab,1947PatternPlayback,HaskinsLab,1950DigitRecognizer,BellLabs,195236DigitPatternTheideawastotrackthefirsttwoformants.
1960-70'sFant,"AcousticTheoryofSpeechProduction",1970BreakthroughinDSPsincethemid1960'–1965FFT–1968HomomorphicProcessing(同态处理)–mid1970'sLinearPredictionAnalysis(线性预测分析)–late1970'sVectorQuantization(矢量量化)Patternmatchingtechniques–1970'sDynamicTimeWarping(动态时间规整)WidelyapplicationofcomputersDARPAstartedSpeechUnderstandingResearch(SUR)programin1970's38Since1980'sSpeechCoding–1980LPC-102.
4kbps–1988FS-10164.
8kbps–1990'sMBE2.
4kbps–ITU-TG-seriesstandard,model-basedVOCODER39Since1980'sSpeechsynthesis–1980Klattcascade/parallelformantsynthesizer–Waveformconcatenationrule-based,TD-PSOLAcorpus-based,unitselection–HMM-basedparametricspeechsynthesis4142第一共振器第二共振器第三共振器第四共振器第五共振器第一共振器第二共振器第二共振器第三共振器第三共振器第四共振器第四共振器第五共振器第五共振器第六共振器++++鼻共振器气管共振器鼻共振器一阶差分滤波脉冲链KLATT声源谱斜率修正L.
F.
声源送气声源擦音噪声源喉声源喉声源串联声道喉声源并联声道(一般不用)擦音噪声源并联声道F0AVOQFLDISQSSTLAHFNPFNZBNPBNZFTPFTZBTPBTZF1B1DF1BF1F2B2F3B3F4B4F5B4CPA2FA3FA4FA5FA6FABANVA1VA2VA3VA4VA5V全通语音输出KlattSynthesizer年份1995年1998年1999年2001年2003年自然度<3.
03.
03.
53.
84.
3STOPWaveformConcatenationSynthesis-iFLYTEKSince1980'sSpeechrecognition–HMM-basedStatisticalpatternrecognitionframework–DevelopmentofVLSIandcomputertechnology–Speechrecognitionsystems1985IBM"Tangora",isolated-wordspeechrecognizer1990IBM"DragonDictate",firstlarge-vocabularyspeech-to-textsystemforgeneral-purposedictation1990'sCMU"Sphinx",continuous-speech,speaker-independentrecognitionsystem1997IBM"ViaVoice"44451997年9月发布Viavoice语音识别软件中文版,从上个世纪70年代开始进行语音技术研究2007-2010年先后发布电话语音搜索,互联网移动语音搜索,GoogleVoiceAction2010年4月收购语音服务提供商Siri,宣布将在iPhone中提供智能语音服务2007年3月以8亿美金价格收购语音搜索业务公司TellMe,加大对语音技术投入2009年10月微软发布WIN7操作系统,集成语音识别技术464748GoogleDuplexGoogleDuplexWhatWeWillBeLearningreviewsomebasicDSPconceptsspeechproductionmodel—acoustics,articulatoryconcepts,speechproductionmodelsspeechperceptionmodel—earmodels,auditorysignalprocessingtimedomainprocessingconcepts—speechproperties,pitch,voiced-unvoiced,energy,autocorrelation,zero-crossingratesshorttimeFourieranalysismethods—digitalfilterbanks,spectrograms,analysis-synthesissystems,vocodershomomorphicspeechprocessing—cepstrum,pitchdetection,formantestimation,homomorphicvocoderlinearpredictivecodingmethods—autocorrelationmethod,covariancemethod,latticemethods,relationtovocaltractmodelsspeechwaveformcodingandsourcemodels—deltamodulation,PCM,mu-law,ADPCM,vectorquantization,multipulsecoding,CELPcodingmethodsforspeechsynthesisandtext-to-speechsystems—physicalmodels,formantmodels,articulatorymodels,concatenativemodelsmethodsforspeechrecognition—theHiddenMarkovModel(HMM)51