噪聲環(huán)境下說話人識(shí)別技術(shù)研究.pdf

1、蘭州理工大學(xué)碩士學(xué)位論文噪聲環(huán)境下說話人識(shí)別技術(shù)研究姓名：曾番申請(qǐng)學(xué)位級(jí)別：碩士專業(yè)：控制理論與控制工程指導(dǎo)教師：陳若珠20080506噪聲環(huán)境卜說話人識(shí)別技術(shù)研究AbstractPreVailingspeal【errecogIIitionsystemsc鋤0btainVeryhigllaccuracyforde姐speech，buttheirperfbnnaI妣willde黟aderapidlyinnoisyenVironmentso

2、wingt0theirmismatchbe細(xì)eentheacousticmodels觚dthetestingspeech，11lerefore，∞iserobusttecIlnologyisacmcialproblemfortheapplicationofspeakcrreC09Ilitionsysteminfeallifc11lepaperfbcus0nt、)Irokeyproblems0fspeakcrrec0印iti衄system

3、，whichcontaillsthefollow：ingaspects：Firstly，thispaperma湖ystudiedMFCC’sextractiontheoryandthcapplicationofwavelet孤“ysist0speechsi印alprocessingWhat’smore，amethodw勰pr01)osedforextractionofanewfeaturepammeterFPBW：whichsubban

4、denergy’scalculation孤dsubb觚denergy’ss印arationbothbased0nwavelet觚alysisTheexperimentsshowedthatthenewpa姍eterFPBWhadbenerrecogIlitionrateandnoisercsistiVepropertyth粗traditionalp猢eterMFCCSecondly，owingt0speechsig艙lcontamina

5、tedbywhitenoise；firstly’Multiscalewavelet觚alysiswasappliedtothep眥essofsigIlalprcproccss，b舔edonit，thetheory0fmultiscale觚alysisw硒appliedtos印aratespeech柚dnOise，卸deIlll孤ccdthespeechconsequentlySeC0ndly，de—noisedspeechi印ut砒嘶g

6、inalspeal【ersystemTheexperimentsshowedthattllismethodelimillatednoiseeffectively鋤dimproVedspeakerreCo伊itionmteinnoisyenVirol加entThirdly，anewrobustspeakerrccogIlitionmethodwasproposedwhichcombinedwaveletdenoisingteclllliq

7、uewiththencwpar鋤eterFPBwneexperimentshowedmatitef￡ectivelyinlprolvedspeakerreC09IlitioninnoisyenViromnentL弱tly，theGaussiallmi】【turemodel(GMM)techniqueswereiIlcre嬲in酉yusedinspeakerrccognitionAimedattheproblemthatwaVelettm

8、sfo咖neededlargecomputationandGMMneededlongtrainingtime，anorthogonalgIlassmixturemodcl(OGMM)w勰employedinorderthatonhogonaltmsfo姍w弱perfo咖edbeforethcuseofexpectationMaximizationalgoritl皿，sothatthecomputationalspeedw嬲improve