Using Parallel WaveGan as vocoder

[Charlottecuc]

Hi. I trained the tacotron2 model and can successfully synthesize voice by using WaveGlow as vocoder. However, when I turned to the Parallel WaveGan (https://github.com/kan-bayashi/ParallelWaveGAN) , the synthzised waveform is quite strange:
.

(In the training time, the hop_size, sample_rate and window_size were set as the same for the tacotron, WaveGlow and waveGan model.)

Previously, I successfully used WaveGan as vocoder to synthesize speech from the FastSpeech acoustic model. The only difference here is that in FastSpeech, mel-spectrogram features were normalized to have zero mean and unit variance before training.

My question is, in your tacotron2 implementation, except audio_norm = audio / self.max_wav_value , is there any other preprocessing of the input mel-spectrogram features? Or, could you kindly give me some advice?

Thank you very much!

[rafaelvalle]

Take a look at this and this

[Charlottecuc]

Thank you. For anyone who tried tacotron2+WaveGan and also encountered this problem, see kan-bayashi/ParallelWaveGAN#169