Neural Vocoding for CycleGAN-Based Voice Conversion

We propose a voice conversion system leveraging recent developments in both voice synthesis and image morphing, which uses CycleGAN to convert mel-spectrograms and neural vocoders to synthesize the converted signals. To evaluate how different vocoders perform in the task, we synthesize converted mel-spectrograms using WaveNet, WaveRNN and MelGAN vocoders. We compare their performances via listening tests, finding that MelGAN and WaveRNN obtained comparable results while WaveNet obtained worse results for converted speech.