Tts: [WIP] tf 2.0 implementation of tts

[erogol]

With the arrival of tf 2.0, I like to give it a try. Probably it would be a separate repo but might be useful for deployment. I'd guess it'd be ready in a week.

[jinfagang]

@erogol How's the work going bro.

[erogol]

https://github.com/erogol/TTS_tf

[erogol]

This is pretty much stuck with TF 2.0 deficiencies especially with changing sequence length causing constant retracing.

[carlfm01]

Hello @erogol, what about https://www.tensorflow.org/guide/ragged_tensor? Were introduced to deal with dimension changing tensor, unfortunately as far as I know there's no benchmark

[erogol]

I don't think I have enough resources right now to work on this. If anyone is volunteering, you are more then welcome to give it a go. I am also ready to help.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions. You might also look our discourse page for further help. https://discourse.mozilla.org/c/tts

[monatis]

hi @erogol, with tf.text and its preprocessing layers returning RaggedTensor, it's much easier to work with inputs of different lengths. Now that I'm working on a neural TTS project in Turkish (novel dataset, sound annotating utility and training code), I'm volunteer to give it a try. What should I do next? Forking the repo, or any other model for collaboration?

[erogol]

You can fork it and give it a go. Then we can iterate together. That would be great .

[monatis]

Great, let you know when I make a progress.

[erogol]

In case anyone is interested. I reiterate TTS with Tensorflow eager mode.

@monatis any updates on your side.

[erogol]

Hello @erogol, what about https://www.tensorflow.org/guide/ragged_tensor? Were introduced to deal with dimension changing tensor, unfortunately as far as I know there's no benchmark

Ragged Tensor looks promising but I am still struggling to convert TTS to TTS_tf due to other TF specific problems.
Such as, in eager mode tf.function tracing, I could not figure out the way to dynamically generate new tensors depending on the input shape here since input tensor has a partial shape at tracing.

[stale[bot]]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions. You might also look our discourse page for further help. https://discourse.mozilla.org/c/tts

[erogol]

I reopen and close this as I added TF 2.0 Tacotron2 implementation.