Adding New Language and Multilingual Training

[yukiarimo]

Hello. I tried out your tutorial for training MeloTTS from scratch (in the docs folder). If I want the model to speak Japanese and English, do I need to train first in English and then in Japanese, or can I do that together? If so, please show me how!

Also, not many languages are supported! How can I add a new language, for example, Russian? I already have a transcribed dataset, and I want to add it to my English+Japanese model. Can you please show me how to do so?

Thank you!

[jidkano]

+1

[jidkano]

@yukiarimo have you found an answer to your question regarding training English first or all languages together?

Also if you don't mind me asking, have you found a way to not start "from scratch" but building on the model's pretrained weights (those shipped with the model package)?

Also interested in whether it's best to provide many audio files with short sentences (2-3 seconds) or fewer, longer audio files (30 seconds to 1 minute)

[yukiarimo]

1. No, I'm still looking for it. Gonna try it tomorrow.
2. Yes, you can. In the shell script (or wherever it is calling the model), just pass your custom model (the official one)
3. Not sure about the maximum length, but I guess it’s around 23 seconds. Still, I would recommend using both short and long samples and hiring a professional voice actress for a more natural sounding (just, don’t use TTS to train a TTS, I know the quality will be decent, but it doesn’t worth it)!

[yukiarimo]

Working on custom TTS. Closing