Some questions

[ZDisket]

1. Are the mel outputs generated compatible with kan-bayashi's ParallelWaveGAN?
2. There's a FastSpeech synthesis example, but not Tacotron2. How to generate speech with the Tacotron2 pretrained model and MelGAN-STFT?

[dathudeptrai]

Hi,

Are the mel outputs generated compatible with kan-bayashi's ParallelWaveGAN?

We have different train/valid split. But generally, the preprocessing steps is the same and the mean/var of our training very close. So let say, i believe it's compatible, you can use Tacotron, Fastpeech generated from this repo and use pretrained models from ParallelWaveGAN. Even it's not compatible, you still can combine by de-norm my mel-spectrogram based on my stats then re-norm based-on kan-bayashi's Parallelwavegan stats :)).

There's a FastSpeech synthesis example, but not Tacotron2. How to generate speech with the Tacotron2 pretrained model and MelGAN-STFT?.

To know how to inference, you can see detail at decode_tacotron2.py or decoder_melgan.py in examples directory. Melgan-STFT is melgan but training with Multi-resolution STFT loss so it's inference same as Melgan original. I will implement AutoModel Class to inference all combinations in the near future :)). Atleast, i will provide google colab soon.

[dathudeptrai]

@ZDisket https://github.com/dathudeptrai/TensorflowTTS/blob/master/notebooks/tacotron2_inference.ipynb. Just in case :D

[ZDisket]

@dathudeptrai
Nice. It seems pretty similar to my notebook.

[dathudeptrai]

@ZDisket great :)). I just uploaded Tacotron pretrained 120K. I'm training multiband melgan, it will 3x faster and improve quality compared with melgan-stft :D.

[ZDisket]

@dathudeptrai Very nice, I have a lot of hope for this repo's Multi-Band MelGAN since I can't get kan-bayashi's to work. It'll be optimal for a user-friendly Windows GUI front end. I'll also retrain my Tacotron2 on the new one.

[dathudeptrai]

why mb-melgan on kan-bayashi not worked ?

[ZDisket]

@dathudeptrai All my predictions had heavy metallic noise, and when it reaches the discriminator train start steps (on another training run) they all become pure noise.

[dathudeptrai]

okay, let see how mb melgan on this repo help you. It will finish training progress on saturday, i think :D.

[ZDisket]

@dathudeptrai
I can't see Tacotron2-120k in the pretrained models section in here. Where is it?

[dathudeptrai]

https://drive.google.com/drive/u/1/folders/1kaPXRdLg9gZrll9KtvH3-feOBMM8sn3_
@ZDisket