A voice conversion method mapping segmented frames with linear multivariate regression

In this paper, we study a different spectral mapping mechanism based on linear multivariate regression (LMR). Such LMR based spectral mapping methods are intended to alleviate the problem of spectral over-smoothing usually encountered by a GMM based method. First, we derive a solution formula to determine the best LMR mapping matrix. Then, for experimental evaluation, we record a parallel corpus, and adopt discrete cepstrum coefficients (DCC) as the spectral features. Next, we label and segment the recorded sentences into the speech units of syllable initials and finals. Hence, an LMR mapping matrix is trained for each syllable initial or final type. In terms of these LMR mapping matrices, we construct a voice conversion system. According to the measured average conversion errors, our system when using the mapping method, LMR_F, can indeed outperform a conventional GMM based voice conversion system. In addition, listening tests are conducted. The results show that the converted speech by our system is slightly better than that converted by a conventional GMM based system.