Singing Voice Conversion Based on Non-Parallel Corpus

Abstract

With the continuous development of deep learning, research on the conversion of singing voice has gradually enriched. The study of singing voice conversion comes from voice conversion. The singing voice conversion is to sing the voice of the target singer without changing the sound content of the source singer. In this paper, we use WORLD vocoder and speech signal processing toolkit (SPTK) to extract the acoustic characteristics of songs and use two mirrored generative adversarial Nets complete the conversion of acoustic features. The experiment realizes the singing conversion of non-parallel corpus. Subjective evaluations show that the songs after the conversion have a good performance in the quality and similarity of the songs.