Some intermediate samples of training vits-2 (non native EN | customer service phone calls dataset)

[p0p4k]

Here is a samples of Voice generated using a custom dataset in LJSpeech Format.

The model ckpts used was G_70000.pth which is so early. Trained on a 2V100 Engine.

Thanks for sharing @athenasaurav ! (#17)

[p0p4k]

You can add more samples as you progress @athenasaurav. Maybe a comparison of another model samples and vits-2 model samples would be great as well later on! Thanks.

[athenasaurav]

Sure @p0p4k

[nmfisher]

Thanks for the great repo, awesome work. Figured I should at least do you the courtesy of sharing some samples (roughly 110k iterations).

https://drive.google.com/drive/folders/13H5qeZ6U-jJVas76ln4TnukT3q1CK6eb?usp=sharing

(If the content is a bit strange, it's because it was all generated via LLaMa 2 in a little demo project I put together
https://youtu.be/GHBOy-duE98).

Chinese will follow soon.

It's still a little unstable in parts (and it does struggle with short phrases), I'm not sure whether that's an artifact of the model, the dataset or the parameters I was training with.

[p0p4k]

Hi, what dataset did you train this on? Thanks for sharing the samples!

[nmfisher]

This is a proprietary multilingual (English/Chinese) dataset, tokenized with ARPABet/pinyin symbols (no word boundaries either, which I think may actually help).

[p0p4k]

Is there a reference sample to compare improvement from VITS-1 ; trained under same conditions.

[athenasaurav]

Yes @nmfisher the audio looks very good. Maybe you show the bad example you mentioned. Also can you give a but specifics about dataset size etc. that will be great.

[yihuitang]

Hi @nmfisher, English sample voices sound very good. Do you have any sample audio for Chinese?