audioedit-models / README.md
mikulskix's picture
Mirror of HT-Demucs ONNX models (4 and 6 stems)
9e3e9f7 verified
|
Raw History Blame Contribute Delete
1.57 kB
---
license: mit
library_name: onnxruntime
pipeline_tag: audio-to-audio
tags:
- onnx
- demucs
- htdemucs
- stem-separation
- source-separation
---
# AudioEdit models (mirror)
Unmodified copies of two ONNX exports of Meta's HT-Demucs (Hybrid Transformer Demucs), used by the
AudioEdit web app, which runs stem separation entirely in the browser (onnxruntime-web).
They are kept here so the app does not depend on third-party repositories staying online.
| File | Stems | Source (exact revision) | SHA-256 |
|---|---|---|---|
| `htdemucs_embedded.onnx` | drums, bass, other, vocals | [timcsy/demucs-web-onnx](https://huggingface.co/timcsy/demucs-web-onnx) @ `92e33df61cfc9eb820272aaa62d2ef6dcf4d950d` | `e5e425c17683f163a472462eb5f5a4ffcd11c31858d57fbd0833b012d8b88077` |
| `htdemucs_6s.onnx` | drums, bass, other, vocals, guitar, piano | [StemSplitio/htdemucs-6s-onnx](https://huggingface.co/StemSplitio/htdemucs-6s-onnx) @ `49df9b6989cf2150840ea65b0bef77a2e471b678` | `48f8e84945579f8ab340e083339e9221e03785dbe733a52c388200b6d3ca779a` |
The files are byte-identical to the sources above. Any processing (for example graph pruning
of the 6-stem model before its in-graph iSTFT) happens in the app at load time.
## License and credits
- Model weights: HT-Demucs by Meta AI Research ([facebookresearch/demucs](https://github.com/facebookresearch/demucs)), MIT License.
- ONNX exports: timcsy (demucs-web) and StemSplitio, both published under the MIT License.
Paper: Rouard, Massa, Défossez, *Hybrid Transformers for Music Source Separation*, ICASSP 2023.