Nonparallel voice conversion based on phoneme classification and eigenvoices

Abstract

This paper proposed a novel algorithm for nonparallel voice conversion based on phoneme classification and eigenvoices. The classification of phoneme, according to phonetic similarity, speaker separability and frequency of occurrence, can avoid the disturbance of linguistic information and spectral smoothing. A speaker adaptation technique of eigenvoices was employed for performing spectral conversion between speakers for each category phoneme, adapting the conversion parameters derived for the pre-stored pairs of speakers to a desired pair, which can relax the parallel constraint effectively. In subjective listening test, an ABX test was performed and the proposed algorithm was preferred to the existing eigenvoice algorithm by 6.44%, and improved quality by 13.14% in terms of mean opinion score (MOS). Both objective and subjective tests demonstrated that the proposed algorithm effectively enhanced speech quality and speaker individuality in a nonparallel manner.