# Third-party notices

- Spleeter: Deezer Research, MIT. The original network source is preserved in `vendor/spleeter/unet.py` with its license at `vendor/spleeter/LICENSE`. The 2-, 4-, and 5-stem ONNX models are exported from the official pretrained checkpoints in [Spleeter v1.4.0](https://github.com/deezer/spleeter/releases/tag/v1.4.0), with each checkpoint's published SHA256 verified before conversion. The export preserves the original weights and model-specific activations, including ELU for 4 / 5 stems and cross-instrument softmax for 5 stems. Model manifests and TensorFlow parity metrics are stored in `public/models/{N}stems/`.
- ONNX Runtime Web: Microsoft, MIT. Runtime package version 1.30.0.
- fflate: Arjun Barrett, MIT. Version 0.8.3, used for batch ZIP exports. Upstream source: https://github.com/101arrowz/fflate .
- `@ffmpeg/ffmpeg`: MIT. `@ffmpeg/core` 0.12.10: GPL-2.0-or-later, as declared by that package. Upstream source and build recipes: https://github.com/ffmpegwasm/ffmpeg.wasm .
- KissFFT WASM: Vilja, MIT. KissFFT itself: BSD-3-Clause. Upstream source: https://github.com/iVilja/kissfft-wasm . The generated factory is exposed as an ESM module by appending one export statement; the FFT implementation is unchanged.
- Example audio `public/samples/demo.mp3` is the official Spleeter repository sample, an excerpt from **Slow Motion Dream** by **Steven M Bryant**, © 2011, featuring **CSoul, Alex Beroza & Robert Siekawitch**, licensed under [Creative Commons Attribution 3.0](https://creativecommons.org/licenses/by/3.0/). The example was copied from https://github.com/deezer/spleeter/blob/master/audio_example.mp3 .

Reference PCM and validation outputs are derived from the same attributed sample solely for the local validation workflow.