TheperformanceofthetraditionalVoiceActivityDetection(VAD)algorithmsdeclinessharplyinlowerSignal-to-NoiseRatio(SNR)environments.Inthispaper,afeatureweightinglikeli-hoodmethodisproposedfornoise-robustVAD.Thecontributionofdynamicfeaturestolikelihoodscorecanbeincreasedviathemethod,whichimprovesconsequentlythenoiserobustnessofVAD.pergencebaseddimensionreductionmethodisproposedforsavingcomputation,whichreducesthesefeaturedimensionswithsmallerpergencevalueatthecostofdegradingtheperformancealittle.ExperimentalresultsonAuroraIIdatabaseshowthatthedetectionperformanceinnoiseenvironmentscanremarkablybeimprovedbytheproposedmethodwhenthemodeltrainedincleandataisusedtodetectspeechendpoints.Usingweightinglikelihoodonthedimension-reducedfeaturesobtainscom-parable,evenbetter,performancecomparedtooriginalfull-dimensionalfeature.
