Add TransformerVC implementation

[ArkhamImp]

The code has been tested, and the model is under training.

[lmxue]

Thanks for your efforts. There are some suggestions for a standard PR, especially for the PR that integrates a new model into Amphion:

• Please format the code using black
• Please detail PR descriptions, including the PR purpose, supported model, etc.
• Please attach the pre-trained models as well as some samples in this PR.
• Please submit a PR from a specific branch instead of the main branch.

[RMSnow]

Thanks @ArkhamImp a lot to integrate our first VC model. For the "New Feature" PR, the review stage is more strict and usually takes several request-change loops.

You can follow these two PRs to provide the inference samples of TransformerVC, so that we can ensure the code is bug-free:

1. DiffWave Vocoder Added #56
2. Add VitsSVC implementation #14

A perfect "New Feature" PR will contain:

• (1) Code;
• (2) Generated Samples;
• (3) Recipe and Doc;
• (4) Pretrained Models in our Huggingface;
• (5) Gradio Demo in our Huggingface.

I think for this PR, you need to provide (1)-(2) at least. Besides, you can schedule an expected date for (3) to inform others. After (1)-(3), we can assign other developers to cooperate with you to accomplish (4) and (5).

[ArkhamImp]

Voice Conversion Samples

Sample 1

Source

121_121726_000020_000001.mp4

Target

672_122797_000002_000002.mp4

Converted

121_121726_000020_000001_672_122797_000002_000002.mp4

Sample 2

Source

1089_134686_000018_000000.mp4

Target

lj001-0003.mp4

Converted

1089_134686_000018_000000_LJ001-0003.mp4

[ArkhamImp]

Samples of VitsVC: https://x8gvg3n7v3.feishu.cn/docx/SPbnd0gHcowovGxCyOLcndfanAb?from=from_copylink

[RMSnow]

Thanks @ArkhamImp ! The VitsVC samples sound better than TransformerVC. You can integrate the both two if you are ready!

[ArkhamImp]

Ready to integrate.