[TTS]Update VITS to support VITS and its voice cloning training on AIShell-3

[HighCWu]

PR types

New features | Bug fixes

PR changes

APIs | Docs

Describe

New features:

• Add voice conversion inference for VITS.
• Add VITS multi speaker and voice cloning training examples for AIShell-3 dataset.
• VITS voice cloning can now use text or audio as input. Due to VITS flow reversibility, it can remove speaker timbre in forward step, and add another speaker timbre in reverse step. For details, see the implementation of the voice conversion. This implementation is referenced from VITS official https://github.com/jaywalnut310/vits/blob/2e561ba58618d021b5b8323d3765880f7e0ecfdb/models.py#L525

Bug fixes:
VITS inference code:
feats_lengths = paddle.to_tensor([paddle.shape(feats)[2]]) -> feats_lengths = paddle.to_tensor(paddle.shape(feats)[2])

Extra:
I added the docs for traning VITS or VITS-VC on aishell-3 dataset, but leave a TODO label for pretrained model. I trained the two new examples with a little modified on aistudio single v100 16GB card with batch_size=24 for 25000 step. This is to verify that training can be performed normally.
I hope the official could release the official version of the pre-trained model trained on 4 cards in the future.

Here are some outputs of my 25000 step models:
test_vits.zip
test_e2e_vits.zip
test_vits_vc.zip
vc_syn_vits_vc_src_text.zip
vc_syn_vits_vc_src_audio.zip

[CLAassistant]

All committers have signed the CLA.

[yt605155624]

Thank you for your contribution. I've been busy publishing the version recently. I'll have a large amount of time to review your code at the end of September.

If you are in Chinese Mainland, you can scan the wechat QR code on the homepage to join our user group and add '子龙' for discussion.

[mergify]

This pull request is now in conflict :(

[yt605155624]

LGTM