---
license: mit
library_name: onnxruntime
tags:
- voice-conversion
- onnx
- vconnx
- rvc
pipeline_tag: audio-to-audio
---

# vconnx · RVC base models (pure ONNX)

Shared base components for [RVC](https://github.com/RVC-Project/Retrieval-based-Voice-Conversion-WebUI)
(MIT) inference in the **vconnx** voice-conversion library: ContentVec-768
content encoder + RMVPE pitch estimator. RVC is **any-to-one** — the target
speaker lives in a separate per-voice `net_g` model (thousands of
community-trained voices exist on HF); these two base models are voice-independent.

| file | role | size |
|---|---|---|
| `contentvec_768l12.onnx` | content encoder (fp32) | 360 MB |
| `contentvec_768l12_q8.onnx` | content encoder (int8) | 91 MB |
| `rmvpe.onnx` | pitch estimator (fp32) | 345 MB |
| `rmvpe_q8.onnx` | pitch estimator (int8) | 94 MB |

ContentVec export parity vs torch: max_abs 8.5e-06. Provenance in `PROVENANCE.md`.

## Usage

```python
from vconnx import VoiceCloner
# reference_voice = an RVC voice MODEL (.onnx path or HF id), not audio
out = VoiceCloner(engine="rvc").clone_voice("source.wav", "owner/repo::voice.onnx")
```