非特定人連續(xù)數(shù)字語音識別研究.pdf

1、大連理工大學(xué)碩士學(xué)位論文非特定人連續(xù)數(shù)字語音識別研究姓名：馮潔申請學(xué)位級別：碩士專業(yè)：計(jì)算機(jī)應(yīng)用技術(shù)指導(dǎo)教師：王健20071201大連理工大學(xué)碩士學(xué)位論文ResearchofSpeakerindependentContinuousChineseDigitStringSpeechRecognitionAbstractSpeechrecognitionhasachievedhighperformanceinlabOurnation’Sres

2、earchonspeechrecognitionstartedin1950’SandisdevelopingrapidlyinrecentyearsandisbeginningtoapplyinpracticalsystemsHoweverinpracticaluse，duetotheexistenceofbackgroundnoise，dialectandtonguespeechrecognitionsystemsarenotused

3、verywidelyEspeciallyforthecascofembeddedsystemswhichareusedincomplexenvironment，itismoreimportanttosolvetheseproblemsSincetheresearchonspeechrecognitionofthelabhasjuststarted，thefoundationoflargevocabularyspeechrecogniti

4、onsystemneedtocreatedictionarywhichrequiresalotoflinguisticknowledgeandalsoneedabigspeechdatabasethepapermainlystudiesspeakerindependentcontinuousChinesedigitstringsspeechrecognition，includingresearchonadaptiveendpointde

5、tectioncontributionofMelfrequencycepstrumcoefficient(MFCC)componentstecognitionrate，choiceofnumbersofHMMstatusandsizeoftrainsetAfterstudyontraditionalendrointdetectionfindthathypothesisofusingafixedCOefficienta(口=1)isnot

6、suitablewhenthesignalnoiseratio(SNR)changesItisnecessarytoaddadaptationprocesswhenthesystembeginstoworkusingsteptostepapproachingmethodTheresultofexperimentshowsthatafteradaptation，thesystemCanapplyinlowerSNRenvironmentM

7、FCCisalleffectivefeatureinspeechrecognitionIntraditionaluse，sincethefirsttwocomponentsofMFCCreflecttheamplitudeofwaveformsandarenegativefortheresultofrecognition，thesetwocomponentsareabandonedHoweverafterexperiments，itis

8、foundthatalthoughthesetwocomponentsdolittlecontributiontodistinguishbetweendigitstheyareusefulindistinguishingbetweenspeechandnoiseSo，thesecomponentscallbeusedinthestageofendpointdetectionExperimentsarealsodoneinthechoic

9、eofhiddenMarkovmodel(HMM)statusnumberandsizeoftrainsetanditiSfoundthatitisreasonabletosetthenumberofH^nⅡstatustobe5andthesizeoftrainsettobe30KeyWords：SpeechRecognition：EndpointDetection；Adaptive；MelFrequencyCepstrumCoeff