Local 6-track separation (browser): model and runtime notice
=============================================================

What runs
- Model: HT-Demucs "htdemucs_6s" (drums, bass, other, vocals, guitar, piano).
- Runtime: ONNX Runtime Web (onnxruntime-web) 1.30.0, WebGPU first, WASM fallback.
- Both run in the visitor's browser. The audio never leaves the device.

Sources, versions and licenses
1. ONNX Runtime Web 1.30.0 (npm: onnxruntime-web, published 2026-09-14)
   License: MIT, Copyright (c) Microsoft Corporation.
   Text: onnxruntime-web-MIT.txt (from https://github.com/microsoft/onnxruntime/blob/main/LICENSE)
   Files served from /vendor/ort/: ort.webgpu.min.mjs, ort-wasm-simd-threaded.asyncify.mjs and
   ort-wasm-simd-threaded.asyncify.wasm (split into two chunks to fit the 25 MiB static asset limit).

2. HT-Demucs weights and original code (Hybrid Transformer Demucs, htdemucs_6s)
   https://github.com/facebookresearch/demucs
   License: MIT, Copyright (c) Meta Platforms, Inc. and affiliates.
   Text: demucs-MIT.txt (https://github.com/facebookresearch/demucs/blob/main/LICENSE).
   Citation: Rouard, Massa, Defossez, "Hybrid Transformers for Music Source Separation", ICASSP 2023.

3. ONNX export of htdemucs_6s
   https://huggingface.co/StemSplitio/htdemucs-6s-onnx (file htdemucs_6s_fp16weights.onnx,
   136428532 bytes, sha256 7ce55792e2231c93fbf92de95f5fd5b3a5e6c89f7db690dfd693e8f1dce56869).
   Model card license: MIT ("matching the original HT-Demucs"). Export tooling:
   https://github.com/StemSplit/demucs-onnx (MIT, Copyright (c) 2026 StemSplit), see demucs-onnx-MIT.txt.

Changes made here (see scripts/demucs-onnx/README.md; the steps are reproducible)
- fold.py: ONNX Runtime constant-folds the float64 inverse-STFT nodes that onnxruntime-web cannot run.
  The result has the same inputs, outputs and numerics.
- shrink.py: stores the large weights as fp16 again (exactly representable, so no precision is lost).
- Served model: /models/htdemucs-6s/ (7 chunks of at most 20 MiB), 146597801 bytes in total,
  sha256 a7fa8997f4f63616b2756a20c92c0000d626e22b8c0dc0d497829eac9d2c73d4.
  The browser reassembles the chunks and verifies this digest before use.

Known limits and open points
- The Demucs authors state that the 6-source model is experimental: the guitar source is acceptable,
  the piano source has a lot of bleed and artifacts.
- The Demucs README says the models were trained on MUSDB HQ plus 800 additional songs. The MIT license
  covers the code and the released weights; it says nothing further about the training data.
- The Meta repository is no longer maintained (its README points to a fork by the original author).
- Model weights license risk was accepted by the project owner on 2026-10-28 and confirmed again on
  2026-10-28 for the production release.