Open-source engines behind the audio tools
Every engine that ships has a written licence here, with the pinned version and the source. audio_tools/licenses.py reads this file at worker boot and refuses to start an engine that has no ## engine: heading. /open-source renders this page for the public.
matchering-2.0.6
- Library: Matchering 2.0.6 by Sergey Grishakov. https://github.com/sergree/matchering · https://pypi.org/project/matchering/2.0.6/
- Licence: GPL-3.0. Server-side use is not "conveying" under GPLv3 §0, so running it for users triggers no source-release obligation. The package is used unmodified; its LICENSE file ships in the worker image.
- What it is: signal processing (RMS, spectrum, peak and stereo-width matching, then a limiter). It is not machine learning and the product never calls it "AI mastering".
- Dependencies: numpy, scipy, soundfile, resampy, statsmodels (BSD-style licences).
openunmix-umxhq
- Library: Open-Unmix (
openunmix) 1.3.0, copyright 2019 Inria, Fabian-Robert Stöter and Antoine Liutkus. https://github.com/sigsep/open-unmix-pytorch · https://pypi.org/project/openunmix/1.3.0/ - Licence: MIT.
- Weights:
umxhq, published under the MIT licence by the authors on Zenodo, record 3370489, DOI https://doi.org/10.5281/zenodo.3370489. Trained on MUSDB18-HQ. Four files, fetched at build time and verified by SHA-256 before every worker boot:
| file | SHA-256 | bytes |
|---|---|---|
| vocals-b62c91ce.pth | b62c91cedbc7a066f1778ead5b5cecb377aa3a46a31af1cce7c5c8769339d083 | 35637796 |
| drums-9619578f.pth | 9619578f885c54737cb0234f9f9a4a679ee4f31438fd77fd1dbe02bb16c2da0a | 35637796 |
| bass-8d85a5bd.pth | 8d85a5bd3f996a8867fca0e8442e077e5a3f5ec747a6112742452a8f347b39c8 | 35637796 |
| other-b52fbbf7.pth | b52fbbf76479e752bd72e02304c602ac7802aa5bfbfb9cd12054b2695d5093ab | 35637796 |
- Excluded on purpose: Open-Unmix
umxl(the library default) is CC BY-NC-SA 4.0 and never runs here; every call namesumxhqexplicitly and a test fails the build if the forbidden name appears. Demucs is not used: its code is MIT but its weight files have no clear commercial licence. - Runtime: PyTorch 2.15.0 (BSD-3), torchaudio 2.11 (BSD-2), CPU build.
Also used
- ffmpeg / ffprobe: invoked as separate binaries (LGPL/GPL depending on the build), never linked.
- RQ (BSD) and Redis for the job queue.