Forward Transformer TTS – a non-autoregressive Transformer for text to speech

as-ideas.github.io

36 pointsdatitran4 comments

We released the Forward version of our Transformer TTS implementation, a text-to-speech Transformer in TensorFlow 2/Keras. Now the model is more robust, fast and controllable.

- Samples: https://as-ideas.github.io/TransformerTTS/

- Code: https://github.com/as-ideas/TransformerTTS

Your work looks good. Code is very clean, well done. What hardware did you use for training? Do you see any potential improvement/application for multispeaker tasks?

Also, what does a publishing house's in-house ML team generally do? Are you building the implementations for peer review replication?

How fast is it?

I cannot speak for the authors, but my impression is that the vocoder (WaveRNN) part of current TTS systems takes much more computing than the part that this research addresses, so this may not help all that much.

Reply on news.ycombinator.com