search

modelscope / FunCodec

FunCodec is a research-oriented toolkit for audio quantization and downstream applications, such as text-to-speech synthesis, music generation et.al.
https://funcodec.github.io/
MIT License
360 stars 30 forks source link

Can the `run.sh` support for the training of Chinese TTS model? #11

Closed Dinxin closed 6 months ago

Dinxin commented 9 months ago

Your README.md said that we can train from scratch using your run.sh script, can it support for Chinese?

ZhihaoDU commented 9 months ago

The run.sh in the egs/LibriTTS/text2speech_laura directory is used to re-produce LauraTTS model on the LibriTTS (English) corpus. In this script, we use g2p_en to tokenize the text to phonemes. Therefore this script is not able to use on Chinese corpora directly. You can replace g2p_en with Chinese g2p packages, such as pypinyin. After preprocessing the text and obtaining the dictionary, this script can be used to train on the Chinese corpora. By the way, the Chinese synthesis model is under internal development and test. We may release it in the future.

Dinxin commented 9 months ago

When will your Chinese model be released approximately?