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

# vconnx · kNN-VC (pure ONNX)

ONNX export of [kNN-VC](https://github.com/bshall/knn-vc) (Baas et al.,
Interspeech 2023, MIT) for the **vconnx** voice-conversion library:
WavLM-Large layer-6 feature extractor + prematched HiFi-GAN vocoder; the
kNN regression step is pure numpy at inference time. Any-to-any voice
conversion at 16 kHz, CPU-only.

| file | role | size |
|---|---|---|
| `wavlm_layer6.onnx` | feature extractor (fp32) | 387 MB |
| `wavlm_layer6_q8.onnx` | feature extractor (int8) | 98 MB |
| `hifigan_knnvc.onnx` | vocoder (fp32) | 63 MB |
| `hifigan_knnvc_q8.onnx` | vocoder (int8) | 25 MB |

Export parity vs the torch reference (max abs):
WavLM see report ·
HiFi-GAN see report
(full reports in the repo).

## Usage

```python
from vconnx import VoiceCloner
out = VoiceCloner(engine="knnvc").clone_voice("source.wav", "reference.wav")
```

Provenance and upstream license: see `PROVENANCE.md`. Weights derive from
[bshall/knn-vc](https://github.com/bshall/knn-vc) (MIT) and
[microsoft WavLM](https://github.com/microsoft/unilm/tree/master/wavlm) (MIT).