NTT123/vietTTS

Prediction time is too slow

GitVHV opened this issue · 2 comments

GitVHV commented

I see an issue coming from text2mel using only 1 cpu and mel2wave being all cpu. Do you have any solution to this problem to optimize processing time? Thank you so much!

Approximetal commented

Reason see this issue jax-ml/jax#2521
I use pytorch version hifigan instead of haiku+jax, it is much faster than previous.
I'm trying to fix text2mel part to solve the recompilation issue, but not finished yet.

  • 👍2
GitVHV commented

Reason see this issue google/jax#2521 I use pytorch version hifigan instead of haiku+jax, it is much faster than previous. I'm trying to fix text2mel part to solve the recompilation issue, but not finished yet.

I also tried running with pytorch but it seems that the input of text2mel doesn't match. Let me know when you get a handle on it. Looking forward to your reply.