The model for Chinese speech doesn't work after training for 1700 epochs about mb-istft-vits2 HOT 3 OPEN

Based on your fork, I've found that you have correct symbols for the cleaner.

But I've also found that you have identical texts for train and validation text. The datasets in validation text should be fewer, and unique from those of training text (Something like 1,2,3,4,5,6,7 for training_text and 8,9,10 for val_text while having a total of 10 datasets.). I suggest split like 5~10 samples from your training text and then pasting it to validation one.

In addition, you have "eval_interval": 100 and "epochs": 1000 and had last update like 13 hours ago. Maybe you meant 1700 steps instead of epochs I guess?

from mb-istft-vits2.

Thanks for your reply! I wonder why the identical texts for training and validation would affect the training outcome. Also, do you think I need to adjust your inference.ipynb when infering Chinese texts? Maybe the inference step goes wrong for Chinese texts (maybe it didn't clean Chinese texts correctly so that the function vcss didn't work). Thank you.

P.S. I did train the model for 1700 epochs since I changed the config.json when training.

from mb-istft-vits2.

Indeed yes. Actually, I forgot to update the notebook version of inference.
You can comment out langdetector like this.

from mb-istft-vits2.