rmvpe-onnx 0.2.3

RMVPE pitch estimator — pure ONNX Runtime inference

rmvpe-onnx

RMVPE pitch estimator with ONNX Runtime backend.

A simple wrapper around ONNX-related code in rvc/lib/rmvpe.py @ 7e03261, RVC-Project/Retrieval-based-Voice-Conversion
Copyright (c) 2023 liujing04, 源文雨, Ftps — MIT License

ONNX model from lj1995/VoiceConversionWebUI
Copyright (c) 2022 lj1995 — MIT License

---

Install

pip

pip install rmvpe-onnx         # Python API only
pip install "rmvpe-onnx[cli]"  # Python API with CLI tool

uv

uv add rmvpe-onnx              # Python API only
uv add "rmvpe-onnx[cli]"       # Python API with CLI tool

[!TIP] This package includes onnxruntime (CPU) as a dependency, which is sufficient for most use cases. For hardware acceleration (CUDA, DirectML, etc.), please install a compatible ONNX Runtime variant. See the ONNX Runtime documentation.

Development

git clone https://github.com/NewComer00/rmvpe-onnx && cd rmvpe-onnx
uv sync --extra dev

# Optional: Run all tests
# uv run tox

CLI

# Optional: Download the ONNX model
# Will be done automatically on first prediction if not done manually
rmvpe-onnx download

# Run pitch prediction with default settings and plot the results
rmvpe-onnx predict audio.wav --plot

For all options, see the CLI Reference or run rmvpe-onnx download --help and rmvpe-onnx predict --help.

Python API

[!NOTE] The Python API returns raw outputs with no confidence thresholding applied. Use confidence to filter frequency yourself if needed.

from rmvpe_onnx import RMVPE
import soundfile as sf

audio, sr = sf.read("audio.wav")
rmvpe = RMVPE()

time, frequency, confidence, activation = rmvpe.predict(audio=audio, sr=sr)

# Optional: Zero out frequency where confidence is below a threshold
# frequency[confidence < 0.03] = 0.0

For full parameter reference and return values, see the API Reference.