Voice Conversion Using Structured Gaussian Mixture Model In Eigen Space

Voice conversion is an important branch in the field of speech signal processing, itstask is to keep the semantic information of voice remains unchanged, only change thespeaker’s personality, and make the personality traits of converted source speech closer tothe target speaker. Traditional voice conversion method most used parallel corpora to jointtraining source and target speaker’s voice model, and then derived the voice conversionfunction. But it is difficult in practice to obtain fully parallel corpora, and union speechmodel training cost a large amount of calculation, more importantly the voice conversionsystem between more speakers is complicated.Under the condition of non-parallel corpora without joint training, a new methodologyof voice conversion in eigen space based on structured Gaussian mixture model isproposed. For every speaker, after the cepstrum feature parameters are extracted, they arefurther mapped to the eigen space which is formed by eigen vectors of scatter matrix of thecepstrum features, then train speaker’s Structured Gaussian Mixture Model in the EigenSpace (SGMM-ES). The source and target speaker’s SGMM-ES are trained respectively，then based on Acoustic Universal Structure (AUS) principle to achieve spectrum transformfunction. Subjective and objective evaluations indicate the conversion performances arequite close to the traditional method under the parallel corpora condition. The results showthe eigen space based on structured Gaussian mixture model for voice conversion underthe non-parallel corpora is effective.This topic research content mainly includes these following aspects:1. Studied the basic principle and mathematical model of speech production, thecharacteristic parameters of voice are analyzed in detail, and skillfully mastering theSTRAIGHT analysis-synthesis platform. 2. Set up a GMM transformed voice platform with joint training under the parallel corpora,and analyzed the problems existing in the traditional voice conversion method.3. Further study of the Acoustic Universal Structure principle, and proposed a StructuredGaussian Mixture Model in the Eigen Space (SGMM-ES).4. Under the condition of non-parallel corpora without joint training, realized a voiceconversion system based on structured Gaussian mixture model in the Eigen Space.5. Using subjective and objective evaluations for the converted speech which obtained byGMM, SGMM, SGMM-ES methods, and analyzed the result in detail to verify theeffectiveness of the proposed method in this paper.