Matcha-TTS icon indicating copy to clipboard operation
Matcha-TTS copied to clipboard

Published 20 hours ago • verified publisher icon shivammehta25

A successfull fa/en implementation report

Open mah92 opened this issue 9 months ago • 4 comments

Hi I wanted to appreciate your work. Thank you. Using this repo, I was able to successfully train two excellent farsi/english tts models. https://huggingface.co/mah92/Khadijah-FA_EN-Matcha-TTS-Model https://huggingface.co/mah92/Musa-FA_EN-Matcha-TTS-Model

By the way, I changed some parts in your repo to do it right, which I shared in the README.md section of the first link.

The most important part was changing the symbols.py part to use multilingual token.txt compatible with sherpa repo...

And here is my issue asking to merge my models in sherpa-onnx: https://github.com/k2-fsa/sherpa-onnx/issues/1779

Thank you again...

mah92 avatar Feb 10 '25 05:02 mah92

I also mentioned that vram usage is proporsional to number of symbols(tokens) used. So we can lower the vram usage by reducing unnecessary symbols(tokens). Am I correct?

mah92 avatar Feb 10 '25 06:02 mah92

Supported in sherpa-onnx!

Please see

https://github.com/k2-fsa/sherpa-onnx/pull/1834

See pre-built APKs at https://k2-fsa.github.io/sherpa/onnx/tts/apk-engine.html

Image

csukuangfj avatar Feb 10 '25 07:02 csukuangfj

https://huggingface.co/mah92/how_to_train_matcha_tts With special changes for mixed persian/english

mah92 avatar Mar 03 '25 04:03 mah92

This is so great to hear! I appreciate you guys experimenting with it :D

shivammehta25 avatar Apr 24 '25 09:04 shivammehta25