资源描述:
《speech acoustics - university of rhode island言语声学-罗德岛大学》由会员上传分享,免费在线阅读,更多相关内容在应用文档-天天文库。
1、SpeechAcousticsProjectOCE471UnderwaterAcousticsJesseHansenSpeechAcousticsProjectJesseHansenAbstract:Inthispaper,basicmethodsforanalyzingrecordedspeecharepresented.ThespectrogramisintroducedandsubsequentlyutilizedinaMatlabenvironmenttorevealpatternsinrecordedvoicedat
2、a.Severalexamplesofspeecharerecorded,analyzed,andcompared.Amodelforvoiceproductionisintroducedinordertoexplainthevarietyoftime-frequencypatternsinthewaveforms.Specifically,asingletubeandthenamulti-tubemodelforthevocaltractareconsideredandrelatedtoresonancesinthespee
3、chspectrum.Itisshownthataseriesofconnectedacoustictubesresultsinresonancessimilartothosethatoccurinspeech.IntroductionMotivation:Considertheproblemofspeechrecognition.Whentwodifferentpeoplespeakthesamephrase(orifonepersonuttersthesamephrasetwice),ahumanlistenerwillg
4、enerallyhavenotroubleunderstandingeachinstanceofthatphrase.Thisleadsustobelievethateventhoughthetwospeakersmayhavedifferentvocalqualities(differentpitch,differentaccents,etc.)theremustbesomesortofinvariantqualitybetweenthetwoinstancesofthespokenphrase.Thinkingaboutt
5、heproblemabitfurther,werealizethatwhentwodifferentpeoplearticulatethesamephrase,theyperformessentiallythesamemechanicalmotions.Inotherwords,theymovetheirmouths,tongue,lips,etc.,intheroughlythesameway.Wehypothesizethatasaresultofthesimilaritiesinspeechmechanicsfrompe
6、rsontopersonthereshouldbesomefeaturesintherecordedspeechwaveformthataresimilarformultipleinstancesofaspokenphrase.Onesuchsetofspeechfeaturesiscalledformants,whichareresonancesinthevocaltract.Thefrequenciesatwhichtheseresonancesoccurareadirectresultoftheparticularcon
7、figurationofthevocaltract.Aswordsarespoken,thespeakermoveshisorhertongue,mouth,andlips,changingtheresonantfrequencieswithtime.Analysisofthesetime-varyingfrequencypatternsformsthebasisforallmodernspeechrecognitionsystems.Organization:Thispaperisbroadlydividedintotwos
8、ections.Part1isconcernedwithanalysisofvoicewaveforms.InPart2,wewilldelveintomodelsforvoiceproductionandrelatethemtothedatapresentedinPart1