基于深度神经网络的说话人信道自适应方法

Inordertohandlethechannelconditiondistortionsbetweentrainandtestspeechinspeakerverification,basedonthedeepneuralnetworks,achanneladaptationapproachwasproposed.First,severalphoneticdeepneuralnetworks(DNNs)weretrainedonthespeechdatasetswithdifferenttypesofchannelconditions.Theacousticfeaturesderivedfromspeakerutteranceswerethenadaptedtoobtaindeepbottleneckfeatures(DBFs)usingtheseDNNs.DBFswereconcatenatedandafeaturedimensionreductionwasperformedusingPCA.Finally,theseDBFsweremodeledbytheidentityvector(i-vector)modelingtechniquewhichisthemostpopularandefficientapproachforspeakerverification.Theachievedi-vectorsfortargetspeakerandtestutteranceswerethenusedtoachievethefinalverificationscores.ResultsontheNISTSRE2010coretestevaluationtaskdemonstratedthatcomparedtothei-vectorbaselinesystem,theproposedapproachiseffectivetoeliminatechanneldistortionsforspeakerverification,andachievessignificantperformanceimprovements.