Hybrid method to convert acoustic features for voice conversion

Abstract

The overly smoothing problem of GMM mapping method is first analyzed, and lost spectral details arising from improper covariance matrixes are considered as the main causation. Thus a hybrid mapping method, which converts envelope-subtracted spectral details by GMM and phone-tied codebook mapping method, is proposed. GMM training in this paper is performed in each phonetic data for faster GMM training. Objective evaluations based on performance indices show that the performance of proposed training method with phonetic information averagely improves 12.87% with tradition GMM training method, and proposed mapping method can improve 27.13% with optimal parameters comparing traditional GMM mapping algorithm with new training method.