A Python implementation for using RVC (Retrieval-based Voice Conversion) via console, Python scripts, or API.

Base code is from: https://github.com/daswer123/rvc-python

I have made a few modifications for my workflow and have removed areas of the readme that I am not using. Please use daswers if you're interested in using this library for other systems or other applications.

1. Create a venv with python 3.11 and activate it

   py -3.11 -m venv venv
   .\venv\Scripts\activate    
   

2. Install necessary requirements. We first install the github repo, then download and install a 3.11 compatible fairseq wheels file (as the OG pypy at 0.12.2 is broken for 3.11), then install pytorch at 2.4.0.

   If you can't use curl for some reason, you can manually download the fairseq wheels file here: https://huggingface.co/Jmica/rvc/resolve/main/fairseq-0.12.4-cp311-cp311-win_amd64.whl?download=true then you can skip the curl command below

   pip install git+https://github.com/JarodMica/rvc-python
   curl -Uri "https://huggingface.co/Jmica/rvc/resolve/main/fairseq-0.12.4-cp311-cp311-win_amd64.whl?download=true" -OutFile "fairseq-0.12.4-cp311-cp311-win_amd64.whl"
   pip install .\fairseq-0.12.4-cp311-cp311-win_amd64.whl
   pip install torch==2.3.1 torchvision==0.18.1 torchaudio==2.3.1 --index-url https://download.pytorch.org/whl/cu121
   

When you first attempt to use this, hubert and rmvpe models will be downloaded from HF into the rvc_python folder in you venv. You can find them there.

Create a new object from RVCInference with the parameters you want set. The defaults are fine, but rvc_models is the root directory that will be looked inside of for RVC models.

Inside root, there should be individual directories named after the desired speaker, and in each, there should be a .pth and .index(if using) file like below.

rvc_models/
├── name_model1/
│   ├── name_model1.pth
│   └── name_model1.index
└── name_model2/
    └── name_model2.pth

An example usage is shown below:

from rvc_python.infer import RVCInference

rvc = RVCInference(models_dir="rvc_models", 
                   device="cuda:0",
                   f0method = "rmvpe",
                   f0up_key=0,
                   index_rate=0.5,
                   filter_radius=3,
                   resample_sr=48000,
                   rms_mix_rate=1,
                   protect=0.33)
rvc.load_model("name_of_rvc_model")
rvc.infer_file("input_path.wav", "output_path.wav")

Thank you to daswer123 for creating the library for easy RVC inference: https://github.com/daswer123/rvc-python