| --- |
| license: mit |
| library_name: onnxruntime |
| pipeline_tag: audio-to-audio |
| tags: |
| - music-source-separation |
| - htdemucs |
| - demucs |
| - onnx |
| - onnxruntime-web |
| - webgpu |
| - stem-separation |
| --- |
| |
| # HTDemucs 6S for ONNX Runtime Web (Audisect build) |
|
|
| Browser-ready ONNX build of **HTDemucs 6S** (Hybrid Transformer Demucs, 6 sources), used by |
| [Audisect](https://audisect.funkz.app), a free stem splitter that runs entirely in the browser. |
|
|
| | File | Size | SHA-256 | |
| |---|---|---| |
| | `htdemucs_6s_web.onnx` | 142,808,361 bytes | `76d177e4cfd7da6aad46e540c5dc929a80f5e839d306a285265f9ec49a28a1ea` | |
|
|
| ## Inputs and outputs |
|
|
| - **Input** `mix`: `float32 [1, 2, 343980]`, stereo, 44.1 kHz, 7.8 s, values in [-1, 1]. |
| - **Output**: `float32 [1, 6, 2, 343980]`, stems in order `drums, bass, other, vocals, guitar, piano`. |
|
|
| Process longer audio in overlapping 7.8 s chunks and crossfade the results. |
|
|
| ## How this file was made |
|
|
| It is derived from [`StemSplitio/htdemucs-6s-onnx`](https://huggingface.co/StemSplitio/htdemucs-6s-onnx) |
| (`htdemucs_6s_fp16weights.onnx`) with Audisect's `scripts/prepare_model.py`: |
|
|
| 1. float64 constants in the iSTFT converted to float32 (ONNX Runtime Web has no float64 `ConstantOfShape`), |
| 2. graph constant-folded with ONNX Runtime's basic optimisations (24,819 to 1,436 nodes), |
| 3. weights stored as float16 with `Cast` nodes. |
|
|
| Checked against the source model on random input: relative RMS error about 3e-5. |
|
|
| ## License and attribution |
|
|
| MIT, same as the upstream weights and export. |
|
|
| - Model: **HTDemucs** by Meta AI Research, [facebookresearch/demucs](https://github.com/facebookresearch/demucs) (MIT). |
| - Paper: Simon Rouard, Francisco Massa, Alexandre Défossez, *Hybrid Transformers for Music Source Separation*, ICASSP 2023. |
| - ONNX export: [StemSplit](https://github.com/StemSplit/demucs-onnx) (MIT). |
|
|
