84基于独立分量分析的语音信号分离算法研究

上海交通大学硕士学位论文基于独立分量分析的语音信号分离算法研究姓名：刘鹏申请学位级别：硕士专业：信号与信息处理指导教师：周军20081201ResearchofSpeechSeparationBasedonIndependentComponentAnalysisABSTRACTSpeechseparationhasbeenahottopicinsignalprocessingsocietyrecentlyyears,whichhasmanyapplicationsandinfluenceintelephoneconference,hearingaid,portabledevices,speechrecognition.Blindsignalprocessingisausefulmethodinspeechseparation,inwhichtheterm“blind”meansthatthesourceitselfandthetransmissionchannelisunknown.Independentcomponentanalysisisthetheoreticalbasisofblindsignalseparation,whichcanbeusedinvarioussignalprocessingfieldsincludingcommunications,image,speech,biology,radar,seismic,sonarandetc.ThethesisstartswiththeresearchofbasictheoriesinBSSandthengoestospeechseparationalgorithmsbasedonconvolutivemodelinfrequencydomain.Themainlyworksarefollowing.Therearemanystudiesininstantaneousmixingmodelintimeandrealdomain,thereforeinrealworldspeechsignalsareconvolvedandwillbetransformedintocomplexvalueinfrequencydomainmethod.Westudyalgorithmsusingthecomplexinformationtoseparatecomplexsignalsandthepermutationprobleminfrequencydomainmethodofconvolutionmodel,proposeanewschemebasedonseparationmatrixinitialization,andhaveaperformancecomparisonlater.Themainmethodofconvolutivespeechseparationinfrequencydomainisthattransformingtheconvolutionintimedomaintomultiplecomputations.Sothecomplextimedomainconvolutivemodelbecomesfrequency-domaininstantaneousmodelsineachbin.Andwecanusethewelldevelopedinstantaneousalgorithmstoestimateseparatingmatrices.Thereforeafterthetransformationbacktotimedomain,wecangetthedeconvolutionFIRfiltersandhavetheestimatedsources.Butthisfrequencydomainmethodbringsinthepermutationproblemwhichcausesperformancedecreaseinseparation.Letusassumeforacouplefrequency-domainsignalsaftertheseparatingineachfrequencybin,wecannotassurethateachoutputchannelonlyconsistsofthecomponentsfromthesamesource.Soifwesimplytransformtheseparatingmatricesintotimedomain,thiswillcausethedeconvolutionfailure.Solvingthepermutationproblemcanbecalledalignment.TraditionalmethodssuchasusingmutualparametersandgeometricDOA(directofarrival)canbeinfluencedbypermutationinlastfrequencybinandthefirststepseparation,thereforehavenotshowngoodresultsinrobustandaccuracyintheschemethatisolatingseparatingstepwithpermutationalignment.ThethesisstudiespermutationalignmentalgorithmsusingcorrelationcoefficientandDOAestimation,implementstheDOAbasedmethod.WealsohaveastudyontheperformanceofDOAmethodanditsinfluenceonthefinalseparatingresults.Weexplorethefundamentallimitationsoftheabovealignmentalgorithmsandthenproposeanewschemebasedonseparationmatrixinitialization,whichisalsoageometricapproach,considertheseparatingsteptogetherwithalignment,cansolvethepermutationwhenseparation.Becauseithasbetterresultsinrobustandaccuracy,thefinalestimatedsourcesarebetterinbothobjectiveandsubjectiveevaluationtargets.Permutationproblemremainsbeingtoughandstillneedsbettersolutions.Thethesisalsointroducesanewfrequencydomainblindseparationschemeusingfrequenciesdependency,whichcanintheoryprobablyavoidthepermutation.Keywords:Speechseparation,independentcomponentanalysis,convolutivemixturesseparationfrequencydomainmethod,ambiguities,permutationalignment1-1[3]................................................................................................................23-1()[1].....................................................183-2[1]..............................................................193-3.......................................................................................................................283-4...................................................................................................................283-5FastICA................................................................................................293-6Infomax................................................................................................293-7JADE......................................................................................................303-84........................................................................................324-1[8]..............................................................................................................334-2[1]......................................................................................354-3m=n=2....................................................................................374-4[31].......................................................................................374-5[43]........................................................................414-6...........................................................................444-716ms..................................................474-8128ms................................................484-9...........................................................................................................484-10.....................................................................................................................504-11...