資源描述：
《改進的核函數(shù)算法及其在說話人辨認中的應用研究》由會員上傳分享，免費在線閱讀，更多相關內(nèi)容在學術論文-天天文庫。
1、北京交通大學碩士學位論文改進的核函數(shù)算法及其在說話人辨認中的應用研究姓名：胡若華申請學位級別：碩士專業(yè)：通信與信息系統(tǒng)指導教師：張有根20080601ABS瞰CTABSTRACT：Duetoitsspecialmeritsofflexibility,economyandaccuracy,speakerrecognitiontechnologyhasabroadapplicationfutureinbiometricsidentityverificationfield．However,speakerrecognitionhassomelimitsinapplicationbecause
2、thetrainingalgorithmiscomplicated,andtherobustnessisnotideal．SupportVectorMachine(SVM)isanewclassificationmethodology．Ithasbeenprovedtobeapowerfultechniqueinpatternclassificationforitsgoodgeneralizationability．ButSVMhassomedisadvantagesinsomeaspectforit’Sstillinthedevelopingstage．Thethesisfocus
3、onhowtoimprovetherecognitionratioandrobustnessofspeakerrecognitionsystembygeneratingnewkernelsbasedonsupervector．Themaincontributionsofthedissertationareasfollows：(1)Theadvancedfeatureparameterextraction．ThisthesisintroducesGaussianMixtureUniversalBackgroundModel(GMM-UBM)intospeakerrecognitionm
4、odeling．UBMUSeSspeaker-independentdistributionparameterstoapproximateparametersforacousticunitswhichareabsentinspecifiedspeaker’Strainingdata．ThenstackthemeansoftheGMMmodelwhichisadaptedbyMAPalgorithmtoformGMMmeansupervector．(2)Theadoptionofnewkernels，suchasKullbackLeiblerDivergencekernel，rinne
5、rproductkernelandNAPkernel．ThesethreenewkernelsareallbasedonGMMsupervector．TheSVMusingthekernelsbasedonGMMsupervectorCanbeusedtoclassifywhollyonthesequence．Meanwhile，inordertoenhancetherobustnessofthesystem，thethesisadoptkernelbasedonNuisanceAttributeProjection．Thiskindofkernelgetsridofredundan
6、tinformationfromthesubspacewhichhasnorelationshipwiththespeakerfeature．(3)Basedonpeoplevoicedatabase，wetakeemulationexperiments．First,wecomparetheadvancedfeatureextractionmethodwithRBFkernelandpolynomialkernel．Second,weapplythesethreenewkernelstospeakeridentification．Fromtheresult，wecanseethese
7、threenewkernelsimprovedrecognitionratioatleastby12％，andNAPkernelimprovedtherobustnessofthesystemalot．KEYWORDS：Speakeridentification；SupportVectorMachine；GMMSupervectorKernel；PrincipalComponentAnalysis(PCA)；NuisanceAttributeProject
