---
license: mit
language:
- en
tags:
- onnx
- audio
- voice-isolation
- speech
- demucs
- bsrnn
pipeline_tag: audio-to-audio
library_name: onnxruntime
---

# VoiceIsolate-Models

Quantized ONNX models used by [VoiceIsolate Pro](https://github.com/Joker5514/VoiceIsolate-Pro) for on-device, GPU-accelerated voice isolation and audio enhancement.

All inference runs 100% client-side in the browser via ONNX Runtime Web (WebGPU / WASM fallback). No server required.

## Models in this Repository

| File | Description | Size | Source |
|---|---|---|---|
| `demucs_v4_quantized.onnx` | Demucs v4 HTDemucs int8-quantized — stem-level voice isolation | ~83 MB | [facebookresearch/demucs](https://github.com/facebookresearch/demucs) |
| `bsrnn_vocals.onnx` | BSRNN Band-Split RNN vocals separator | ~45 MB | [crlandsc/bsrnn](https://github.com/crlandsc/bsrnn) |

## Usage

These models are fetched automatically by `ml-worker-fetch-cache.js` in VoiceIsolate Pro. They are cached in IndexedDB after the first download and never re-fetched.

```js
// MODEL_REGISTRY entry in ml-worker-fetch-cache.js
demucs_v4: {
  path: 'models/demucs_v4_quantized.onnx',
  sizeBytes: 87_031_808,
  cdnUrls: ['https://huggingface.co/Joker5514/VoiceIsolate-Models/resolve/main/demucs_v4_quantized.onnx']
},
bsrnn_vocals: {
  path: 'models/bsrnn_vocals.onnx',
  sizeBytes: 3_870_554,
  cdnUrls: ['https://huggingface.co/Joker5514/VoiceIsolate-Models/resolve/main/bsrnn_vocals.onnx']
}
```

## License

MIT. Model weights inherit the licenses of their respective upstream projects:
- Demucs: MIT
- BSRNN: MIT