Synthesized audio quality is poor by using pre-trained WaveNet vocoder model

[mindmapper15]

I trained tacotron 2 model with LJSpeech Dataset with default setting.

Synthesized mel spectrogram seems to fine but when I converted synthesized mel spectrogram to audio with pre-trained WaveNet Vocoder model (ljspeech.wavenet.mol.v2) described in README page, the synthesized audio has very poor quality compared to synthesized mel spectrogram.

I extracted synthesized mel spectrogram from feats.ark and save those with npy format.

This is the synthesized mel spectrogram from LJSpeech dev set.

And this is the converted audio from synthesized mel spectrogram I attatched

LJ049-0151.zip

I've done peak normalize to the synthesized audio because its amplitude is too small

Is this model trained properly? If so, did I put synthesized mel spectrogram to vocoder in wrong way?

[kan-bayashi]

It seems something wrong.
Please check the following points:

• Are the feature setting the same among the models? (fmax, fmin, ...)
• Did you use normalized features? (not denorm)

[kan-bayashi]

For reference (our pretrained model sample):
https://drive.google.com/drive/folders/1JFNZapygWsHiP2CXMjTraLzf98h-tEBF?usp=sharing

[mindmapper15]

• Are the feature setting the same among the models? (fmax, fmin, ...)

I used ljspeech.wavenet.mol.v2 model which was trained with mel range 80 ~ 7600Hz
This mel range is same as default setting in LJSpeech TTS recipe in espnet.

• Did you use normalized features? (not denorm)

I extracted mel spectrogram from egs/ljspeech/tts1/exp/char_train_no_dev_pytorch_train_pytorch_tacotron2/outputs_model.last1.avg.best_decode_denorm/feats.ark. As far as I know, this feature file contains synthesized mel spectrogram which is denormalized at stage 5 in run.sh by following command.

apply-cmvn --norm-vars=true --reverse=true data/${train_set}/cmvn.ark \ scp:${outdir}/${name}/feats.scp \ ark,scp:${outdir}_denorm/${name}/feats.ark,${outdir}_denorm/${name}/feats.scp

so probably it's not the normalized feature I think.

[kan-bayashi]

Yes, that is de-normalized feature.
Please use outputs_model.last1.avg.best_decode/feats.ark as the input of WaveNet.

[mindmapper15]

LJ049-0151.zip

I changed wavenet input to normalized feature and now it seems work!
Thank you for the corrections!

Problem solved so I'll close this issue.