Add VitsSVC implementation

[viewfinder-annn]

Add implementation of VITS-based model for Singing Voice Conversion task

[RMSnow]

Two additional general comments:

1. Add Amphion's copyright information for all the newly added files.
2. Add some comments (descriptions and detail instructions) for some functions.

[RMSnow]

Is this file the same as egs/_template/run.sh? You can create a soft link to it. It is easy for our future repair.

[viewfinder-annn]

Actually not 🥲 vits module needs special Cython module initialization other than common run.sh for svc task:
https://github.com/viewfinder-annn/AmphionPublic/blob/main/egs/svc/VitsSVC/run.sh#L9

So is it feasible to split Cython module initialization into another .sh file? In this case we can indicate this in readme file and reuse egs/_template/run.sh.

[RMSnow]

Use Black for formatting your code. See this blog.

[RMSnow]

Is this class same as

Amphion/models/tts/vits/vits.py

Line 151 in 42ca016

class SynthesizerTrn(nn.Module):

We need to merge them into a common one.

[viewfinder-annn]

No, the internal encoding is different, one for text and acoustic condition the other, which affects forward/infer functions either. So in my opinion it can not be merged.

[RMSnow]

Is this file necessary?

[viewfinder-annn]

Yes, as discussed before, whisper extractor needs modules/whisper_extractor/assets/mel_filters.npz to extract feature properly, while this file is initially ignored by the universal *.npz configuration. Adding this line can make this file preserved by git version control.

[zhizhengwu]

Any audio samples to support this PR?

[viewfinder-annn]

@zhizhengwu
Any audio samples to support this PR?

Here are two samples which are converted from M4Singer dataset to Opencpop dataset, with original sample and SoVits4.1 model's output.
VitsSVC model uses ContentVec and Whisper feature and hifigan as generator, is trained from scratch for 110k steps.
SoVits4.1 model uses Whisper feature and nsf-hifigan as generator, is fine-tuned on a 330k steps-pretrained base model for 110k steps.

	Tenor-3 -> opencpop_female1	Alto-6 -> opencpop_female1
Original	Tenor-3	Alto-6
SoVits 4.1	Tenor-3_SoVits4.1_opencpop_female1	Alto-6_SoVits4.1_opencpop_female1
VitsSVC	Tenor-3_VitsSVC_opencpop_female1	Alto-6_VitsSVC_opencpop_female1

[RMSnow]

Modify the modules.vits. Let @lmxue know it.