so-vits-svc-pipeline

A python CLI pipeline script for running UVR tasks, So-VITS-SVC model inference and ffmpeg merge with end-to-end automation.

1. Prerequisite

Before running pipeline script, the following software must be installed first:

1. UltimateVocalRemoverGUI v5.6.0
2. FFMpeg
3. Latest version of So-VITS-SVC 4.1 (Chinese document)

1.1 UVR GUI

Follow the Manual Windows Installation instruction in UltimateVocalRemoverGUI. If anaconda is used, try the following command:

conda create -n uvr python=3.9
conda activate uvr
pip install -r requirements.txt

After installation, run UVR GUI and download required models (optional) via: python UVR.py

Note: The root path of UVR GUI is denoted as uvr_path.

1.2 FFMpeg

1. Download and extract FFMpeg (both ffmpeg and ffprobe are required)
2. Two choices:
   • Add the executable files into PATH environment, this script will detect them automatically. Check this step by opening a terminal (cmd) and type ffmpeg -version
   • Specify the path of ffmpeg and ffprobe in the ffmpeg section of the config json manually, or pass by argument --ffmpeg.ffmpeg_path <ffmpeg_executable> --ffmpeg.ffprobe_path <ffprobe_executable>

1.3 So-VITS-SVC 4.1

Here uses a community-maintain variation of So-VITS-SVC 4.1 for simplifying preprocessing and training. Download and extract all files, the root path of So-VITS-SVC is denoted as vits_path later.

2. How to use

This repository contains two main modules to run the pipeline: UVR-http-service and pipeline-client. Each of them is deployed to different directory.

2.1 UVR-http-service

Put uvr_api.py and uvr_api_server.py into uvr_path (the root directory where UVR GUI is extracted). Then run python uvr_api_server.py instead of the previous command python UVR.py.

This time, UVR GUI is able to accept tasks through HTTP protocol (listening on port 8090 by default) and perform the corresponding automation steps (setting up parameters, inputs and outputs).

2.2 pipeline-client

To run So-VITS-SVC pipeline, ensure run_svc_new.py and uvr_api_client.py are in the same directory first, then run python run_svc_new.py [-c vits_config.json] [-p default] [-k] to perform end-to-end automation.

Arguments for this pipeline:

• -c or --config: config file for SO-VITS-SVC pipeline, default: vits_config.json
• -p or --profile: SO-VITS-SVC inference arguments (defined in the vits section of the config file), default: default
• -k or --keep: keep intermediate files during this pipeline (these files will be cleaned up after finish)

Other arguments will be set into config json, for example: --file.input <input_file> --file.output <output_file_or_dir> will overwrite the input file and output file in the config json

There are 3 preset config files provided in this repository, the corresponding process pipeline are also attached:

1. vits_config.json (default)
   • just separate vocal-instrumental tracks and remove harmony for the vocal track
   • relatively fast and can provide acceptable quality
   • 
2. vits_config_with_harmony.json
   • reserve harmony track and run VITS separately for the harmony track
   • slower and the F0 prediction for harmony track may fail quite often
   • can produce better quality for the songs with simple harmony
   • 
3. vits_config_multi_spk.json (experimental)
   • for multi-speaker VITS models only
   • 1 speaker per channel is recommended (one for L channel and another for R channel)
   •