SoftVC VITS Singing Voice Conversion

Model Overview

A singing voice coversion (SVC) model, using the SoftVC encoder to extract features from the input audio, sent into VITS along with the F0 to replace the original input to acheive a voice conversion effect. Additionally, changing the vocoder to NSF HiFiGAN to fix the issue with unwanted staccato.

Notice

• This fork has some modifications to make it work better on Windows and with smaller multi-speaker datasets.
• Some modifications have also been made to pitch inference for better performance.
• There is one gui using PySide6 inference_gui.py and one gui using PyQt5 currently a work in progress inference_gui2.py
  • Inference GUI 2 features experimental TalkNet integration, in-program recording, as well as other features like timestretching with rubberband. Instructions can be found below under Inference GUI 2 header.
  • Also check out GothicAnon's GUI, written in tkinter. See here, under So-Vits-SVC for more info.

• The current branch is the 32kHz version, which requires less vram during inferencing, as well as faster inferencing speeds, and datasets for said branch take up less disk space. Thus the 32 kHz branch is recommended for use.
• If you want to train 48 kHz variant models, switch to the main branch. NOTE: effusiveperiscope does not maintain a 48khz branch.

Colab notebook scripts

Colab training notebook (EN)

Colab training notebook (CN)

Colab inference notebook

Inference

Instructions for CLI inference can be found in the original repository or here. Note that CLI inference is very clunky and generally not recommended for production. Instructions for inference using various GUIs are available below.

Required downloads

• Download softVC hubert model:hubert-soft-0d54a1f4.pt
  • Place under hubert.
• Download pretrained models G_0.pth and D_0.pth
  • Place under logs/32k.
  • Pretrained models are required, because from experiments, training from scratch can be rather unpredictable to say the least, and training with a pretrained model can greatly improve training speeds.
  • The pretrained model includes云灏, 即霜, 辉宇·星AI, 派蒙, and 绫地宁宁, covering the common ranges of both male and female voices, and so it can be seen as a rather universal pretrained model.
  • The pretrained model exludes the optimizer speaker_embedding section, rendering it only usable for pretraining and incapable of inferencing with.

The following shell commands can be used for downloading with bash.

# For simple downloading.
# hubert
wget -P hubert/ https://github.com/bshall/hubert/releases/download/v0.1/hubert-soft-0d54a1f4.pt
# G&D pretrained models
wget -P logs/32k/ https://huggingface.co/innnky/sovits_pretrained/resolve/main/G_0.pth
wget -P logs/32k/ https://huggingface.co/innnky/sovits_pretrained/resolve/main/D_0.pth

Inference GUI 2

For Inference GUI 2, you need to pip install PyQt5. Additional features may be available based on other dependencies:

• OPTIONAL - You PROBABLY DO NOT NEED THIS: For timestretching support, you need to install BOTH the rubberband standalone program, ensuring the rubberband executable is on your PATH, and the python module pip install pyrubberband. Note that installing pyrubberband installs PySoundFile which needs to be uninstalled, and SoundFile will need to be reinstalled.
• OPTIONAL - For TalkNet support, you need to pip install requests and also install this ControllableTalkNet fork.

Basic Usage

Models should be placed in separate folders within a folder called models, in the same directory as inference_gui2.py by default. Specifically, the file structure should be:

so-vits-svc-eff\
	models\
		TwilightSparkle
			G_*****.pth
			D_*****.pth
			config.json

If the proper libraries are installed, the GUI can be run simply by running inference_gui2.py. If everything goes well you should see something like this:

All basic workflow occurs under the leftmost UI panel.

1. Select a speaker based on the listed names under Speaker:.
2. Drag and drop reference audio files to be converted onto Files to Convert. Alternatively, click on Files to Convert to open a file dialog.
3. Set desired transpose (for m2f vocal conversion this is usually 12 i.e. an octave, or leave it 0 if the reference audio is female) under Transpose.
4. Click Convert. The resulting file should appear under results.

The right UI panel allows for recording audio directly into the GUI for quick fixes and tests. Simply select the proper audio device and click Record to begin recording. Recordings will automatically be saved to a recordings folder. The resulting recording can be transferred to the so-vits-svc panel by pressing Push last output to so-vits-svc.

Common issues

• When converting: TypeError: Invalid file: WindowsPath('...') Ensure that PySoundFile is not installed (pip show pysoundfile). PySoundFile is a deprecated version of SoundFile. After uninstalling pysoundfile, run pip install soundfile==0.10.3.post1 --force-reinstall
• When trying to run with TalkNet: Couldn't parse TalkNet response. Ensure that you are running alt_server.py in the TalkNet fork.

Running with TalkNet

For TalkNet support, you need to pip install requests and also install this ControllableTalkNet fork. Instead of running talknet_offline.py, run alt_server.py (if you use a batch script or conda environment to run TalkNet, you should use it to run alt_server.py). This will start a server that can interface with Inference GUI 2. The TalkNet server should be started before Inference GUI 2.

Next, starting Inference GUI 2 should show a UI like this:

The rightmost panel shows controls for TalkNet which are similar to those used in the web interface. Some items special to this interface:

• There is currently no "Custom model" option. To add additional models you should modify the model jsons in Controllable TalkNet.
• Recordings can be also be transferred from the recording panel to the TalkNet panel.
• Files can be provided under Provide input audio through clicking for a file dialog or drag-and-drop.
• In order to push output from TalkNet through so-vits-svc, check Push TalkNet output to so-vits-svc. For production work it is advised to first try one generation without this box checked to see if there are artifacts in the TalkNet output. The output will use the speaker selected in the leftmost panel.

Updates

According to incomplete statistics, it seems that training with multiple speakers may lead to worsened leaking of voice timbre. It is not recommended to train models with more than 5 speakers. The current suggestion is to try to train models with only a single speaker if you want to achieve a voice timbre that is more similar to the target. Fixed the issue with unwanted staccato, improving audio quality by a decent amount.
The 2.0 version has been moved to the 2.0 branch.
Version 3.0 uses the code structure of FreeVC, which isn't compatible with older versions.
Compared to DiffSVC , diffsvc performs much better when the training data is of extremely high quality, but this repository may perform better on datasets with lower quality. Additionally, this repository is much faster in terms of inference speed compared to diffsvc.

Training

Dataset preparation

All that is required is that the data be put under the dataset_raw folder in the structure format provided below.

dataset_raw
├───speaker0
│   ├───xxx1-xxx1.wav
│   ├───...
│   └───Lxx-0xx8.wav
└───speaker1
    ├───xx2-0xxx2.wav
    ├───...
    └───xxx7-xxx007.wav

Data pre-processing.

1. Resample to 32khz

python resample.py

2. Automatically sort out training set, validation set, test set, and automatically generate configuration files.

python preprocess_flist_config.py
# Notice.
# The n_speakers value in the config will be set automatically according to the amount of speakers in the dataset.
# To reserve space for additionally added speakers in the dataset, the n_speakers value will be be set to twice the actual amount.
# If you want even more space for adding more data, you can edit the n_speakers value in the config after runing this step.
# This can not be changed after training starts.

3. Generate hubert and F0 features/

python preprocess_hubert_f0.py

After running the step above, the dataset folder will contain all the pre-processed data, you can delete the dataset_raw folder after that.

Training.

python train.py -c configs/config.json -m 32k

Inferencing.

Use inference_main.py

• Edit model_path to your newest checkpoint.
• Place the input audio under the raw folder.
• Change clean_names to the output file name.
• Use trans to edit the pitch shifting amount (semitones).
• Change spk_list to the speaker name.