STATISTICALEIGENVOICE:SPEAKERFEATURESWITHINS+NFRAMEWORK AND A WAY TOWARDS LANGUAGE-INDEPENDENTVOICECONVERSION
JunxunYin FengHuang · 2005
Thispaper presents astatistical method forspeaker feature extraction andvoice conversion within Sinusoidal+Noise (S+N)modeling framework. Withfundamental researches onspeaker characteristics embedded intheparameter sets ofS+Nmodel, wefound thevector sets ofStatistical EigenVoice(SEV)andWeighted Statistical EigenVoice (wSEV), whicharebasis vectors ofGMM representation, havesignificant properties: approximately speaker-dependent and language-independent. Piered bythefeature vectors ofSEV andwSEV,we address anewalgorithm forcontext-free voice conversion. Subjective tests suggest that theSEVbasedmethodachieves convincing results while maintaininghighsynthesis quality incomparison tothetraditional LPCapproaches.