摘要
InordertoimprovetheperformanceofdeceptiondetectionbasedonChinesespeechsignals,amethodofsparsedecompositiononspectralfeatureisproposed.First,thewaveletpackettransformisappliedtodividethespeechsignalintomultiplesub-bands.Bandcepstralfeaturesofwaveletpacketsareobtainedbyoperatingthediscretecosinetransformonlogarithmicenergyofeachsub-band.ThecepstralfeatureisgeneratedbycombingMelFrequencyCepstralCoefficientandWaveletPacketBandCepstralCoefficient.Second,K-singularvaluedecompositionalgorithmisemployedtoachievethetrainingofanover-completemixturedictionarybasedonboththetruthanddeceptivefeaturesets,andanorthogonalmatchingpursuitalgorithmisusedforsparsecodingaccordingtothemixturedictionarytogetsparsefeature.Finally,recognitionexperimentsareperformedwithvariousclassifiedmodules.Experimentalresultsshowthatthesparsedecompositionmethodhasbetterperformancecompariedwithconventionaldimensionreducedmethods.Therecognitionaccuracyofthemethodproposedinthispaperis78.34%,whichishigherthanmethodsusingotherfeatures,improvingtherecognitionabilityofdeceptiondetectionsystemsignificantly.
出版日期
2019年01月11日(中国Betway体育网页登陆平台首次上网日期,不代表论文的发表时间)