Singing Voice Conversion Based on WD-GAN Algorithm
Wei Zhao, Weikang Wang, Yuanyuan Sun, Tang Tang · 2019
The research of singing voice conversion(SVC) has attracted more and more attention in the field of artificial intelligence. This paper proposes a WD-GAN algorithm for singing voice conversion. WORLD vocoder is used to extract the feature tensor of singing. The feature tensor and the singer's timbre label c are sent into G model for training. Using the differentiability of Wasserstein distance description, the problem of GAN training instability is solved. Through the game process of D model and G model, we can train the timbre characteristics of many singers and realize many-to-many singing voice conversion. Both objective and subjective evaluation experiments show that WD-GAN algorithm has better conversion effect than baseline network without Wasserstein distance in SVC.