Spaces:

Snow-White-995

/

Bert-VITS2-Shanghainese

like 9

Sleeping

First of all, thank you for making this project. I was trying to modify the official bert vits 2 project but the structure is so messy... And I found this project. Your modification follows the vits base code and more understandable.

I tried to archive Cantonese ability by:

• Replace zaonhe.json to jyutjyu.json (with .ocd2 also) in text/shanghainese.py
• Replace sh_symbols array in text/symbols.py with config from CjangCjengh:

https://sjtueducn-my.sharepoint.com/:u:/g/personal/cjang_cjengh_sjtu_edu_cn/EbtTeDZwyyxEgDO8h1CwSXoBBiCNAnFwvkNEhHxVtjRBpQ?e=xtr3sx

I assume after removal of the marks, the array should be ["#", "N", "a", "b", "d", "e", "f", "g", "h", "i", "j", "k", "l", "m", "n", "o", "p", "r", "s", "t", "u", "v", "w", "x", "y", "z", "\u00e6", "\u00e7", "\u00f8", "\u014b", "\u0153", "\u0235", "\u0250", "\u0251", "\u0252", "\u0253", "\u0254", "\u0255", "\u0257", "\u0258", "\u0259", "\u025a", "\u025b", "\u025c", "\u0263", "\u0264", "\u0266", "\u026a", "\u026d", "\u026f", "\u0275", "\u0277", "\u0278", "\u027b", "\u027e", "\u027f", "\u0282", "\u0285", "\u028a", "\u028b", "\u028c", "\u028f", "\u0291", "\u0294", "\u02a6", "\u02ae", "\u02b0", "\u02b7", "\u02c0", "\u02d0", "\u02e5", "\u02e6", "\u02e7", "\u02e8", "\u02e9", "\u0303", "\u031a", "\u0325", "\u0329", "\u1d00", "\u1d07", "\u2191", "\u2193", "\u2205", "\u2c7c", " "]

And.... It did not work. The audio sounds like a people with stroke.

Could you help me if you know where goes wrong? And do I have to retrain the model? Also how? Currently the training script in this repo is missed.