FACTORIALSCALEDHIDDENMARKOVMODELFOR
POLYPHONICAUDIOREPRESENTATIONANDSOURCESEPARATION
edricF´evotte2andMauriceCharbit1AlexeyOzerov1,*,C´
1
2
InstitutTelecom,TelecomParisTech,CNRSLTCICNRSLTCI,TelecomParisTech
37-39,rueDareau,75014Paris,France
{alexey.ozerov,cedric.fevotte,maurice.charbit}@telecom-paristech.fr
ABSTRACT
WepresentanewprobabilisticmodelforpolyphonicaudiotermedFactorialScaledHiddenMarkovModel(FS-HMM),whichgener-alizesseveralexistingmodels,notablytheGaussianscaledmix-turemodelandtheItakura-SaitoNonnegativeMatrixFactorization(NMF)model.Wedescribetwoexpectation-maximization(EM)algorithmsformaximumlikelihoodestimation,whichdifferbythechoiceofcompletedataset.ThesecondEMalgorithm,basedonareducedcompletedatasetandmultiplicativeupdatesinspiredfromNMFmethodology,exhibitsmuchfasterconvergence.WeconsidertheFS-HMMindifferentcon gurationsforthedif cultproblemofspeech/musicseparationfromasinglechannelandreportsatisfyingresults.
IndexTerms—FactorialhiddenMarkovmodel,Gaus-sianscaledmixturemodels,nonnegativematrixfactorization,expectation-maximizationalgorithm,audiosourceseparation.
1.INTRODUCTION
Manyadvancedprobabilisticmodelsforpolyphonicaudiohavebeenproposedforsourceseparation[1,2,3,4]andsemanticin-formationretrieval(e.g.,musictranscription[5]).Thereisacloserelationbetweensourceseparationandsemanticinformationre-trievalinthesensethattheoutcomeofonetaskcanhelptheother.Forexample,singingvoiceseparationisusedtoimprovesingeridenti cationin[6]and,conversely,mainmelodylineestimationisusedtoimprovesourceseparationin[7].Therefore,itseemspromisingtoperformthesetasksjointly(e.g.,modelsfrom[3]and[5]allowjointsourceseparationandmusictranscription),thusex-ploitingtheinformationfrombothacousticandsemanticlevelsinanoptimalway.Suchmodelsshouldbegenerativesoastoallowsourcereconstructionandincludeasemanticlevel,expressedforexampleintermsofsomehiddenstates.
Inthisworkwepresentanewgenericmodelwiththeabove-mentionedproperties.WeconsiderthefamilyofmodelswheretheShort-TimeFourierTransform(STFT)oftheaudiosignalistakenasrealizationsofzero-meanproperGaussianmultivariaterandomvariables(possiblyconditionallyonsomehiddenstatevariables)[1,2,4,8].Suchmodelshaveprovenef cientforsourcesepara-tion;theyallowinparticularstraightforwardsourceestimationvia(adaptive)Wiener ltering.
Benaroyaetal.[2]modeleachsourceSTFTbyaGaussianMixtureModel(GMM)modulatedbyaframe-dependentampli-tudeparameteraccountingfornonstationarity,leadingtotheGaus-*
A.OzerovisnowwithIRISA,MetissGroup,Rennes,France.
ThisworkwassupportedinpartbytheFrenchANRprojectSARAH.
978-1-4244-3679-8/09/$25.00 ©2009 IEEE121
sianScaledMixtureModel(GSMM).Anothermodel,bettersuitedtopolyphony,isproposedin[1]andwaslinkedtoNonnegativeMatrixFactorization(NMF)withtheItakura-Saito(IS)divergencebyF´evotteetal.[8].ThismodeltakeseachsourceSTFTtobeasumofseveralelementaryGaussiancomponentswith xedspectrumamplitude-modulatedinframe.Inshort,theGSMMim-plicitlyassumesthesourcetobemonophonicwithmanypossiblestateswhiletheIS-NMFmodelassumesthesourcetobepoly-phonic,i.e.,asum1
ofmanyelementarycomponentswithsimplespectralsignature.Theaimofthispaperistobridgebetweenthetwomod-elsandintroduceageneralhybridmodelthatencompassesboththeGSMMandNMFmodel.Moreoverweincorporatetime-persistenceinthemodelthroughMarkovmodeling.Inessence,ourmodelassumestheobservedcompositesignaltobeasumofindependentcomponentseachmodeledaHiddenMarkovModel(HMM)withdiagonalcovariancematricesamplitude-modulatedinframe.Ourmodel,termedFactorialScaledHiddenMarkovModel(FS-HMM),isdescribedindetailsinSection2.ThenwedescribeinSection3twoExpectation-Maximization(EM)algo-rithmsforMLestimation.InSection4theconvergencespeedofthealgorithmsiscompared,andtheFS-HMMisappliedindif-ferentcon gurationstotheproblemofspeech/musicseparationfromasinglechannel.Finally,conclusionsaredrawninSection5.2.FACTORIALSCALEDHIDDENMARKOVMODELLetxn=[xfn]Ff=1bethecomplex-valuedSTFTofsomeaudiodata,termedobservation(f=1,...,Fisafrequencybinindex,nmodel=1,...isconsidered:
,Nisatimeframeindex).ThefollowingcompositexK
n=
k=1
ck,n,(1)
wherethecomplex-valuedvectorsck,n=[ck,fn]Fascomponentsandassumedmutuallyindependentf=1arereferredto(i.e.,overk).LetIk,nbeJk-statesdiscreterandomvariables(r.v.)independentoverk.Weassumeforeachcomponentck,nthefollowingmixturemodel
c Jk,n=k
i=1
uki,n1(Ik,n=i),(2)
wherethesub-componentsuki,n=[uk,stateiandki,fn]Ftimef=1areassumedin-dependent(overcomponentn)propercomplex
1In
otherwords,withGSMMasummationtakesplaceintheproba-bilitydensityfunction(pdf)ofthesourceSTFTwhilewithIS-NMFthe
summationtakesplaceintheSTFTdomain.
[21] A. Ozerov et al., “Factorial scaled hidden Markov model for polyphonic audio representation and source separation,” in Proc. WASPAA 2009. [22]...
[21] A. Ozerov et al., “Factorial scaled hidden Markov model for polyphonic audio representation and source separation,” in Proc. WASPAA 2009. [22]...
热门文档