Wasserstein Generative Adversarial Networks for Singing Voice Conversion


Junchen Lua, Kun Zhoua, Berrak Sismanb,a, Haizhou Lia
junchen@u.nus.edu   zhoukun@u.nus.edu   berraksisman@u.nus.edu   haizhou.li@nus.edu.sg

a Dept. of Electrical and Computer Engineering, National University of Singapore
b Information Systems Technology and Design, Singapore University of Technology and Design

----------------------------> Model Architecture <-----------------------



Fig.1 The training phase of the proposed F0-VAW-GAN SVC framework. Blue boxes are involved in the training.


Fig.2 The run-time conversion phase of the proposed F0-VAW-GAN SVC framework. Red boxes have been trained during the training phase.

Experimental Setup:

VAW-GAN (SID): VAW-GAN system that converts spectrum (with no conditioning on decoder), F0 is converted through LG-based linear transformation;
VAW-GAN (SID + F0) (proposed): Converts the spectrum with VAW-GAN conditioned on LG-based F0, where F0 is converted with LG-based linear transformation;
The codes of this research are available here.


-----------------------> Singing Samples <-----------------------



Source VAW-GAN (SID) VAW-GAN (SID + F0) (Proposed) Target
Male-to-Male
Male-to-Female