RVC-ONNX

RVC-ONNX is a lightweight inference wrapper around ONNX Runtime for running Retrieval-based Voice Conversion (RVC) models.
This project supports RVC models and performs inference in FP16 for faster execution and reduced memory usage.

Features

• RVC model inference with ONNX Runtime
• FP16 inference for performance
• F0 extraction with RMVPE
• Command-line interface for quick usage

Quickstart

This project uses uv for packaging and execution.

Clone the repository:

git clone https://github.com/dev6699/rvc-onnx.git
cd rvc-onnx
uv sync

Get models from: here

• vec-256-layer-9.onnx (for V1)
• vec-768-layer-12.onnx (for V2)
• rmvpe.onnx

Usage:

# For V1 model
uv run rvc-onnx --input assets/input_1.wav --output assets/output_1.wav --model-path models/v1.onnx --vec-path models/vec-256-layer-9.onnx

# For V2 model
uv run rvc-onnx --input assets/input_1.wav --output assets/output_1.wav --model-path models/v2.onnx --vec-path models/vec-768-layer-12.onnx

# Help
uv run rvc-onnx --help
usage: rvc-onnx [-h] --input INPUT --output OUTPUT [--sid SID] [--f0-up-key F0_UP_KEY]
                [--model-path MODEL_PATH] [--vec-path VEC_PATH] [--rmvpe-path RMVPE_PATH]

Run RVC inference with ONNX models

options:
  -h, --help            show this help message and exit
  --input INPUT         Path to input audio file
  --output OUTPUT       Path to save output audio file
  --sid SID             Speaker ID
  --f0-up-key F0_UP_KEY
                        Pitch shift amount
  --model-path MODEL_PATH
                        Path to rvc ONNX model
  --vec-path VEC_PATH   Path to ContentVec model
  --rmvpe-path RMVPE_PATH
                        Path to RMVPE model

Result

Input 1	Output 1
input_1.webm	output_1.webm

---

Input 2	Output 2
input_2.webm	output_2.webm

License

This project is licensed under the terms of the MIT license.