lora-svc

Singing Voice Conversion based on Whisper & neural source-filter BigVGAN

Black technology based on the three giants of artificial intelligence:

OpenAI's whisper, 680,000 hours in multiple languages

Nvidia's bigvgan, anti-aliasing for speech generation

Microsoft's adapter, high-efficiency for fine-tuning

LoRA is not fully implemented in this project, but it can be found here: LoRA TTS & paper

use pretrain model to fine tune

https://user-images.githubusercontent.com/16432329/231021007-6e34cbb4-e256-491d-8ab6-5ce4e822da21.mp4

Dataset preparation

Necessary pre-processing:

• 1 accompaniment separation, UVR
• 2 cut audio, less than 30 seconds for whisper, slicer

then put the dataset into the data_raw directory according to the following file structure

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

Install dependencies

• 1 software dependency

  pip install -r requirements.txt

• 2 download the Timbre Encoder: Speaker-Encoder by @mueller91, put best_model.pth.tar into speaker_pretrain/

• 3 download whisper model multiple language medium model, Make sure to download medium.pt，put it into whisper_pretrain/

  Tip: whisper is built-in, do not install it additionally, it will conflict and report an error

• 4 download pretrain model maxgan_pretrain_32K.pth, and do test

  python svc_inference.py --config configs/maxgan.yaml --model maxgan_pretrain_32K.pth --spk ./configs/singers/singer0001.npy --wave test.wav

Data preprocessing

use this command if you want to automate this:

python3 prepare/easyprocess.py

or step by step, as follows:

• 1， re-sampling

  generate audio with a sampling rate of 16000Hz

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

  generate audio with a sampling rate of 32000Hz

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

• 2， use 16K audio to extract pitch

  python prepare/preprocess_f0.py -w data_svc/waves-16k/ -p data_svc/pitch

• 3， use 16K audio to extract ppg

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

• 4， use 16k audio to extract timbre code

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

• 5， extract the singer code for inference

  python prepare/preprocess_speaker_ave.py data_svc/speaker/ data_svc/singer

• 6， use 32k audio to generate training index

  python prepare/preprocess_train.py

• 7， training file debugging

  python prepare/preprocess_zzz.py -c configs/maxgan.yaml

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
└── 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

Train

• 0， if fine-tuning based on the pre-trained model, you need to download the pre-trained model: maxgan_pretrain_32K.pth

  set pretrain: "./maxgan_pretrain_32K.pth" in configs/maxgan.yaml，and adjust the learning rate appropriately, eg 1e-5

• 1， start training

  python svc_trainer.py -c configs/maxgan.yaml -n svc

• 2， resume training

  python svc_trainer.py -c configs/maxgan.yaml -n svc -p chkpt/svc/***.pth

• 3， view log

  tensorboard --logdir logs/

Inference

use this command if you want a GUI that does all the commands below:

python3 svc_gui.py

or step by step, as follows:

• 1， export inference model

  python svc_export.py --config configs/maxgan.yaml --checkpoint_path chkpt/svc/***.pt

• 2， use whisper to extract content encoding, without using one-click reasoning, in order to reduce GPU memory usage

  python whisper/inference.py -w test.wav -p test.ppg.npy

• 3， extract the F0 parameter to the csv text format

  python pitch/inference.py -w test.wav -p test.csv

• 4， specify parameters and infer

  python svc_inference.py --config configs/maxgan.yaml --model maxgan_g.pth --spk ./data_svc/singers/your_singer.npy --wave test.wav --ppg test.ppg.npy --pit test.csv

  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 --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

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

• 5, post by vad

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

Source of code and References

Adapter-Based Extension of Multi-Speaker Text-to-Speech Model for New Speakers

AdaSpeech: Adaptive Text to Speech for Custom Voice

https://github.com/nii-yamagishilab/project-NN-Pytorch-scripts/tree/master/project/01-nsf

https://github.com/mindslab-ai/univnet [paper]

https://github.com/openai/whisper/ [paper]

https://github.com/NVIDIA/BigVGAN [paper]