AnewmethodologyofvoiceconversionincepstrumeigenspacebasedonstructuredGaussianmixturemodelisproposedfornon-parallelcorporawithoutjointtraining.Foreachspeaker,thecepstrumfeaturesofspeechareextracted,andmappedtotheeigenspacewhichisformedbyeigenvectorsofitsscattermatrix,therebytheStructuredGaussianMixtureModelintheEigenSpace(SGMM-ES)istrained.Thesourceandtargetspeaker’sSGMM-ESarematchedbasedonAcousticUniversalStructure(AUS)principletoachievespectrumtransformfunction.Experimentalresultsshowthespeakeridentificationrateofconversionspeechachieves95.25%,andthevalueofaveragecepstrumdistortionis1.25whichis0.8%and7.3%higherthantheperformanceofSGMMmethodrespectively.ABXandMOSevaluationsindicatetheconversionperformanceisquiteclosetothetraditionalmethodundertheparallelcorporacondition.TheresultsshowtheeigenspacebasedstructuredGaussianmixturemodelforvoiceconversionunderthenon-parallelcorporaiseffective.
