復(fù)雜信道下的說(shuō)話人識(shí)別.pdf

1、博士學(xué)位論文論文題目復(fù)雜信道1=的說(shuō)話人識(shí)別作者姓名學(xué)科專業(yè)導(dǎo)師姓名完成時(shí)間王仁華教授二oo七年八月爹～上囂、一舞。攀。叁一蘭拶拜掣酮渺i中～中國(guó)科學(xué)技術(shù)人學(xué)博fj學(xué)位論義摘舞AbstractTmsthesisfocusesonthesessionvariabilitytext—independentspeakerrecognitionInthisthesisWeexplorethepossibilitiestoobtainthehig

2、hidentificationratewhilemaintainingthelittlecomputationload，Theyareexploredin：l、frontfeatureextraction；2)testalgorithmofGMM_UBM；3)SVMbasedsystem；4)sessionvariabilityreductionNovelandefficientalgorithmsareproposedIncludin

3、g：ThefixedframelengthacousticfeatureisalwaysadoptedinspeakerrecognitionThevoicedandunvoicedsoundistreatedequallyButtheunvoicedsoundismorelikeawhitenoisesignal，thevoicedsoundreflectsthemovementofthevocaltractandisaperiodi

4、csignal，SOmorespeakerinformationiscontainedinthevoicedsoundThevariableframelengthfeatureextractionprocedureisproposedFurthermorethefeatureofthevoicedsoundhasmoreweightthantheunvoicedsoundinmodeltrainingprocedurenleEERoft

5、hedynamicfeatureCallbereducedby10％againstthetraditionalfixedframelengthfeatureTheGMM_UBMisthestate—oftheartsysteminspeakerrecognitionThelog—likelihoodtestalgorithmisadoptedinthetestprocedureInthisthesis，theangleofthemode

6、ldivergenceisintroducedtoreplacetheloglikelihoodinthetestprocedureTheangleofthemodeldivergenceCanacquirealmostthesarneresultasthelog—likelihoodFurthermore，thescoresoftwosystemcarlbefused，theEERofthefusedsystemCalldecreas

7、eby12％15％againsteithersystemInrecentyearstheSVMhasmadeencouragingprogressinspeakerrecognitionThreemethodsarediscussedtoimprovetheSVMbasedsystemnTheoptimizedGMMmeansupervectorandwei【ghtsupervectoraresuggestedastheinputoft

8、heSVM111esuggestedsupervectorcanoutperformthetraditionalGMMmeallsupervectorby20％inEER21ThemodeldivergeneeandtheangleofthemodeldivergencearealsoprososedastheinputoftheSVM，andtheyCanbecombinedwiththeGLDStoimpmveperformance

9、3)Duetotheinsufficiencyofdatafromtargetspeakers，theSVMalgorithmfrequentlyencounterstheclassimbalanceproblemwhichmayintroducesevereperformancedegradationTwostrategiesareproposedtoselectimpostor’ssamplesinSVMtrainingThefir