152334H / DL-Art-School

TorToiSe fine-tuning with DLAS
GNU Affero General Public License v3.0
214 stars 96 forks source link

Possible way to avoide English accent while training on other languages #52

Open andreibezborodov opened 1 year ago

andreibezborodov commented 1 year ago

Hello!

Thanks for great work! I was trying to finetune a model using non English datasets (Russian, etc.). The resulting voice is really good, but I keep getting the result with super strong English accent even after long training. Are there any possible ways to reduce the accent (or ideally get rid of it)? I guess that the problem is because of the fine-tuning process using English model..

HobisPL commented 1 year ago

You can try this. https://github.com/152334H/DL-Art-School/discussions/51

andreibezborodov commented 1 year ago

You can try this. #51

Thank you! I would also mention that for training on Cyrillic letters it is also required to change the english_cleaners to basic_cleaners. I've made a new tokenizer and started training, but the results so far are not good.

Can you please tell how big was your dataset and for how long did you train? I wonder how big should be a dasatet for fine-tuning on a new language.

pivolan commented 11 months ago

@andreibezborodov hi, can you help me with start finetuning on another languages? @cherpekat telegram. Cannot connect with you by email in your github profile.