Variational Inference with adversarial learning for end-to-end Singing Voice Conversion based on VITS

[![Hugging Face Spaces](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Spaces-blue)](https://huggingface.co/spaces/maxmax20160403/sovits5.0) GitHub Repo stars

[中文文档](./README_ZH.md) The tree [bigvgan-mix-v2](https://github.com/PlayVoice/whisper-vits-svc/tree/bigvgan-mix-v2) has good audio quality The tree [RoFormer-HiFTNet](https://github.com/PlayVoice/whisper-vits-svc/tree/RoFormer-HiFTNet) has fast infer speed No More Upgrade

This project targets deep learning beginners, basic knowledge of Python and PyTorch are the prerequisites for this project;
This project aims to help deep learning beginners get rid of boring pure theoretical learning, and master the basic knowledge of deep learning by combining it with practices;
This project does not support real-time voice converting; (need to replace whisper if real-time voice converting is what you are looking for)
This project will not develop one-click packages for other purposes;

vits-5.0-frame

A minimum VRAM requirement of 6GB for training
Support for multiple speakers
Create unique speakers through speaker mixing
It can even convert voices with light accompaniment
You can edit F0 using Excel

https://github.com/PlayVoice/so-vits-svc-5.0/assets/16432329/6a09805e-ab93-47fe-9a14-9cbc1e0e7c3a

Model properties

Feature	From	Status	Function
whisper	OpenAI	✅	strong noise immunity
bigvgan	NVIDA	✅	alias and snake	The formant is clearer and the sound quality is obviously improved
natural speech	Microsoft	✅	reduce mispronunciation
neural source-filter	Xin Wang	✅	solve the problem of audio F0 discontinuity
pitch quantization	Xin Wang	✅	quantize the F0 for embedding
speaker encoder	Google	✅	Timbre Encoding and Clustering
GRL for speaker	Ubisoft	✅	Preventing Encoder Leakage Timbre
SNAC	Samsung	✅	One Shot Clone of VITS
SCLN	Microsoft	✅	Improve Clone
Diffusion	HuaWei	✅	Improve sound quality
PPG perturbation	this project	✅	Improved noise immunity and de-timbre
HuBERT perturbation	this project	✅	Improved noise immunity and de-timbre
VAE perturbation	this project	✅	Improve sound quality
MIX encoder	this project	✅	Improve conversion stability
USP infer	this project	✅	Improve conversion stability
HiFTNet	Columbia University	✅	NSF-iSTFTNet for speed up
RoFormer	Zhuiyi Technology	✅	Rotary Positional Embeddings

due to the use of data perturbation, it takes longer to train than other projects.

USP : Unvoice and Silence with Pitch when infer vits_svc_usp

Why mix

mix_frame

Plug-In-Diffusion

plug-in-diffusion

Setup Environment

Install PyTorch.
Install project dependencies
```
pip install -i https://pypi.tuna.tsinghua.edu.cn/simple -r requirements.txt
```
Note: whisper is already built-in, do not install it again otherwise it will cuase conflict and error
Download the Timbre Encoder: Speaker-Encoder by @mueller91, put best_model.pth.tar into speaker_pretrain/.
Download whisper model whisper-large-v2. Make sure to download large-v2.pt，put it into whisper_pretrain/.
Download hubert_soft model，put hubert-soft-0d54a1f4.pt into hubert_pretrain/.
Download pitch extractor crepe full，put full.pth into crepe/assets.

Note: crepe full.pth is 84.9 MB, not 6kb

Download pretrain model sovits5.0.pretrain.pth, and put it into vits_pretrain/.

python svc_inference.py --config configs/base.yaml --model ./vits_pretrain/sovits5.0.pretrain.pth --spk ./configs/singers/singer0001.npy --wave test.wav

Dataset preparation

Necessary pre-processing:

Separate voice and accompaniment with UVR (skip if no accompaniment)
Cut audio input to shorter length with slicer, whisper takes input less than 30 seconds.
Manually check generated audio input, remove inputs shorter than 2 seconds or with obivous noise.
Adjust loudness if necessary, recommend Adobe Audiiton.

Put the dataset into the dataset_raw directory following the structure below.

dataset_raw
├───speaker0
│   ├───000001.wav
│   ├───...
│   └───000xxx.wav
└───speaker1
├───000001.wav
├───...
└───000xxx.wav

Data preprocessing

python svc_preprocessing.py -t 2

-t: threading, max number should not exceed CPU core count, usually 2 is enough. After preprocessing you will get an output with following structure.

data_svc/
└── waves-16k
│    └── speaker0
│    │      ├── 000001.wav
│    │      └── 000xxx.wav
│    └── speaker1
│           ├── 000001.wav
│           └── 000xxx.wav
└── waves-32k
│    └── speaker0
│    │      ├── 000001.wav
│    │      └── 000xxx.wav
│    └── speaker1
│           ├── 000001.wav
│           └── 000xxx.wav
└── pitch
│    └── speaker0
│    │      ├── 000001.pit.npy
│    │      └── 000xxx.pit.npy
│    └── speaker1
│           ├── 000001.pit.npy
│           └── 000xxx.pit.npy
└── hubert
│    └── speaker0
│    │      ├── 000001.vec.npy
│    │      └── 000xxx.vec.npy
│    └── speaker1
│           ├── 000001.vec.npy
│           └── 000xxx.vec.npy
└── whisper
│    └── speaker0
│    │      ├── 000001.ppg.npy
│    │      └── 000xxx.ppg.npy
│    └── speaker1
│           ├── 000001.ppg.npy
│           └── 000xxx.ppg.npy
└── speaker
│    └── speaker0
│    │      ├── 000001.spk.npy
│    │      └── 000xxx.spk.npy
│    └── speaker1
│           ├── 000001.spk.npy
│           └── 000xxx.spk.npy
└── singer
│   ├── speaker0.spk.npy
│   └── speaker1.spk.npy
|
└── indexes
    ├── speaker0
    │   ├── some_prefix_hubert.index
    │   └── some_prefix_whisper.index
    └── speaker1
        ├── hubert.index
        └── whisper.index

Re-sampling

Generate audio with a sampling rate of 16000Hz in ./data_svc/waves-16k

python prepare/preprocess_a.py -w ./dataset_raw -o ./data_svc/waves-16k -s 16000

Generate audio with a sampling rate of 32000Hz in ./data_svc/waves-32k

python prepare/preprocess_a.py -w ./dataset_raw -o ./data_svc/waves-32k -s 32000

Use 16K audio to extract pitch

python prepare/preprocess_crepe.py -w data_svc/waves-16k/ -p data_svc/pitch

Use 16K audio to extract ppg

python prepare/preprocess_ppg.py -w data_svc/waves-16k/ -p data_svc/whisper

Use 16K audio to extract hubert

python prepare/preprocess_hubert.py -w data_svc/waves-16k/ -v data_svc/hubert

Use 16k audio to extract timbre code

python prepare/preprocess_speaker.py data_svc/waves-16k/ data_svc/speaker

Extract the average value of the timbre code for inference; it can also replace a single audio timbre in generating the training index, and use it as the unified timbre of the speaker for training
```
python prepare/preprocess_speaker_ave.py data_svc/speaker/ data_svc/singer
```

Use 32k audio to extract the linear spectrum

python prepare/preprocess_spec.py -w data_svc/waves-32k/ -s data_svc/specs

Use 32k audio to generate training index
```
python prepare/preprocess_train.py
```
Training file debugging
```
python prepare/preprocess_zzz.py
```

Train

If fine-tuning is based on the pre-trained model, you need to download the pre-trained model: sovits5.0.pretrain.pth. Put pretrained model under project root, change this line
```
pretrain: "./vits_pretrain/sovits5.0.pretrain.pth"
```
in configs/base.yaml，and adjust the learning rate appropriately, eg 5e-5.

batch_size: for GPU with 6G VRAM, 6 is the recommended value, 8 will work but step speed will be much slower.

Start training

python svc_trainer.py -c configs/base.yaml -n sovits5.0

Resume training

python svc_trainer.py -c configs/base.yaml -n sovits5.0 -p chkpt/sovits5.0/sovits5.0_***.pt

Log visualization
```
tensorboard --logdir logs/
```

sovits5 0_base

sovits_spec

Inference

Export inference model: text encoder, Flow network, Decoder network

python svc_export.py --config configs/base.yaml --checkpoint_path chkpt/sovits5.0/***.pt

Inference

if there is no need to adjust f0, just run the following command.

python svc_inference.py --config configs/base.yaml --model sovits5.0.pth --spk ./data_svc/singer/your_singer.spk.npy --wave test.wav --shift 0

if f0 will be adjusted manually, follow the steps:
1. use whisper to extract content encoding, generate test.vec.npy.
```
python whisper/inference.py -w test.wav -p test.ppg.npy
```
2. use hubert to extract content vector, without using one-click reasoning, in order to reduce GPU memory usage
```
python hubert/inference.py -w test.wav -v test.vec.npy
```
3. extract the F0 parameter to the csv text format, open the csv file in Excel, and manually modify the wrong F0 according to Audition or SonicVisualiser
```
python pitch/inference.py -w test.wav -p test.csv
```
4. final inference
```
python svc_inference.py --config configs/base.yaml --model sovits5.0.pth --spk ./data_svc/singer/your_singer.spk.npy --wave test.wav --ppg test.ppg.npy --vec test.vec.npy --pit test.csv --shift 0
```

Notes
- when --ppg is specified, when the same audio is reasoned multiple times, it can avoid repeated extraction of audio content codes; if it is not specified, it will be automatically extracted;
- when --vec is specified, when the same audio is reasoned multiple times, it can avoid repeated extraction of audio content codes; if it is not specified, it will be automatically extracted;
- when --pit is specified, the manually tuned F0 parameter can be loaded; if not specified, it will be automatically extracted;
- generate files in the current directory:svc_out.wav
Arguments ref

args --config --model --spk --wave --ppg --vec --pit --shift

name config path model path speaker wave input wave ppg wave hubert wave pitch pitch shift

args	--config	--model	--spk	--wave	--ppg	--vec	--pit	--shift
name	config path	model path	speaker	wave input	wave ppg	wave hubert	wave pitch	pitch shift

post by vad

python svc_inference_post.py --ref test.wav --svc svc_out.wav --out svc_out_post.wav

Train Feature Retrieval Index (Optional)

To increase the stability of the generated timbre, you can use the method described in the Retrieval-based-Voice-Conversion repository. This method consists of 2 steps:

Training the retrieval index on hubert and whisper features Run training with default settings:

python svc_train_retrieval.py

If the number of vectors is more than 200_000 they will be compressed to 10_000 using the MiniBatchKMeans algorithm. You can change these settings using command line options:

usage: crate faiss indexes for feature retrieval [-h] [--debug] [--prefix PREFIX] [--speakers SPEAKERS [SPEAKERS ...]] [--compress-features-after COMPRESS_FEATURES_AFTER]
                                                 [--n-clusters N_CLUSTERS] [--n-parallel N_PARALLEL]

options:
  -h, --help            show this help message and exit
  --debug
  --prefix PREFIX       add prefix to index filename
  --speakers SPEAKERS [SPEAKERS ...]
                        speaker names to create an index. By default all speakers are from data_svc
  --compress-features-after COMPRESS_FEATURES_AFTER
                        If the number of features is greater than the value compress feature vectors using MiniBatchKMeans.
  --n-clusters N_CLUSTERS
                        Number of centroids to which features will be compressed
  --n-parallel N_PARALLEL
                        Nuber of parallel job of MinibatchKmeans. Default is cpus-1

Compression of training vectors can speed up index inference, but reduces the quality of the retrieve. Use vector count compression if you really have a lot of them.

The resulting indexes will be stored in the "indexes" folder as:

data_svc
...
└── indexes
    ├── speaker0
    │   ├── some_prefix_hubert.index
    │   └── some_prefix_whisper.index
    └── speaker1
        ├── hubert.index
        └── whisper.index

At the inference stage adding the n closest features in a certain proportion of the vits model Enable Feature Retrieval with settings:
```
python svc_inference.py --config configs/base.yaml --model sovits5.0.pth --spk ./data_svc/singer/your_singer.spk.npy --wave test.wav --shift 0 \
--enable-retrieval \
--retrieval-ratio 0.5 \
--n-retrieval-vectors 3
```
For a better retrieval effect, you can try to cycle through different parameters: --retrieval-ratio and --n-retrieval-vectors

If you have multiple sets of indexes, you can specify a specific set via the parameter: --retrieval-index-prefix

You can explicitly specify the paths to the hubert and whisper indexes using the parameters: --hubert-index-path and --whisper-index-path

Create singer

named by pure coincidence：average -> ave -> eva，eve(eva) represents conception and reproduction

python svc_eva.py

eva_conf = {
    './configs/singers/singer0022.npy': 0,
    './configs/singers/singer0030.npy': 0,
    './configs/singers/singer0047.npy': 0.5,
    './configs/singers/singer0051.npy': 0.5,
}

the generated singer file will be eva.spk.npy.

Data set

Name	URL
KiSing	http://shijt.site/index.php/2021/05/16/kising-the-first-open-source-mandarin-singing-voice-synthesis-corpus/
PopCS	https://github.com/MoonInTheRiver/DiffSinger/blob/master/resources/apply_form.md
opencpop	https://wenet.org.cn/opencpop/download/
Multi-Singer	https://github.com/Multi-Singer/Multi-Singer.github.io
M4Singer	https://github.com/M4Singer/M4Singer/blob/master/apply_form.md
CSD	https://zenodo.org/record/4785016#.YxqrTbaOMU4
KSS	https://www.kaggle.com/datasets/bryanpark/korean-single-speaker-speech-dataset
JVS MuSic	https://sites.google.com/site/shinnosuketakamichi/research-topics/jvs_music
PJS	https://sites.google.com/site/shinnosuketakamichi/research-topics/pjs_corpus
JUST Song	https://sites.google.com/site/shinnosuketakamichi/publication/jsut-song
MUSDB18	https://sigsep.github.io/datasets/musdb.html#musdb18-compressed-stems
DSD100	https://sigsep.github.io/datasets/dsd100.html
Aishell-3	http://www.aishelltech.com/aishell_3
VCTK	https://datashare.ed.ac.uk/handle/10283/2651
Korean Songs	http://urisori.co.kr/urisori-en/doku.php/