Audio parameters for different Vocoder and TTS models

[arif334]

I've been training different models (from the recipe) with my Bangla (Bengali) dataset containing 12 hours of clean speech. So far, I've trained GlowTTS, Hifi-GAN, MB-MelGAN, and WaveRNN. They are working fine with each other.

But, then I trained Tacotron2-DDC and Tacotron2-DCA. They generate a somewhat good voice with the default GL vocoder but do not work with the trained neural vocoders (Hifi-GAN, MB-MelGAN). Combining Tacotron2 with the neural vocoders generates just noise.

I've noticed some differences in the audio parameters of Tacotron2 and the other models.

In Tacotron2 config, I've found these:

signal_norm=false
log_func=np.log
mel_fmax=8000
spec_gain=1.0

While in the other models, these values are as follows:

signal_norm=true
log_func=np.log10
mel_fmax=null
spec_gain=20

My question is, what should I do to be able to use Tacotron2 with other models? Should I retrain the Tacotron2 (and/or a vocoder) model with matching parameters? Changing those values to the trained model's config files (in either Tacotron2 or vocoder models) generates robotic (but intelligible) sound BTW.

[erogol]

Audio parameters should match

[arif334]

Audio parameters should match

Thanks @erogol. I've got that. Now I have to retrain one of the models. Which of the following two options would you suggest?

1. Retrain Tacotron2 with matching parameters of a vocoder.
2. Retrain a vocoder with matching parameters of Tacotron2.

[erogol]

Matching vocoder would be easier