Meta AI (Demucs) · Utility

HTDemucs stem separation

Non-commercial onlyMIT (code) / research-only (weights)

A separator: it takes a sound you already have and splits it into stems. Run it from a sound's “Split into stems” action; it takes no prompt. The code licence and the weights licence differ here, and the shipped catalog gave the code licence for both.

Demucs upstream project audio.cpp

The licence

Summarized from the linked source documents, last checked 25 August 2026. Every claim links to its source; anything unconfirmed is marked unconfirmed.

MIT (code) / research-only (weights)

Non-commercial only

The Demucs repository is MIT, and that is the licence most sources give, including audio.cpp's conversion table and the shipped catalog. The MIT licence covers the code. Demucs' maintainer has stated that it does not cover the pretrained weights, which are provided for scientific purposes only. You download and run the weights, so this page treats the model as non-commercial.

Conditions

Maintainer statement, verbatim: "The model weights are not covered by the MIT license, and are provided only for scientific purposes."
The upstream README is silent on the weights, which is why the MIT label spread so widely. Third-party mirrors relabel the weights MIT; that relabelling is not backed by any statement from the rights-holder that could be found.

What about the audio you generate?

Neither the licence nor the README says anything about separated output. Rights in the stems you extract mostly follow your rights in the input audio, not this model's licence. Separating a track you have no rights to gives you no rights in the stems.

Correction to the shipped catalog: The shipped catalog says "MIT (unverified)" with commercial use "yes". That is wrong. MIT is the licence on the code, and the maintainer has said explicitly that it does not extend to the weights. The common claim that the v4 weights are CC-BY-NC-4.0 could not be confirmed either: Demucs v1 was, but no current upstream page states it for HTDemucs. The confirmed constraint is "scientific purposes only", so this is shown as non-commercial, not conditional.

Based on the licence text as of the date below. Licences change; the links carry the authoritative version.

Specs

Transcribed from the shipped catalog, audiocpp_models_catalog.json — that file is authoritative if a number here disagrees with it.

Vendor: Meta AI (Demucs)
Catalog id: audiocpp-htdemucs-sep
audio.cpp loader family: htdemucs
Can do: Separation
Weights repo: audio-cpp/audio.cpp-gguf
Weights file: HTDemucs-GGUF/htdemucs-q8_0.gguf
Account needed to download: No
Max clip length: 600s
Sample rate: 44.1kHz stereo
Approx. VRAM: 3.0 GB
In the picker by default: No — opt in from Settings
Run end-to-end here: No — catalogued, not tested here

Getting it

There is no per-model install. One native binary runs all of these, so getting a model is a download of its weights and nothing else — no virtualenv, no torch pin, no CUDA wheel to match against your driver.

From the CLI

soundbox model install audiocpp
soundbox pull audiocpp-htdemucs-sep

The binary is fetched from audio.cpp's GitHub releases the first time you ask for it and cached under your own data directory. Soundbox ships no native binaries and vendors none.
Weights land in <models dir>/audiocpp/audiocpp-htdemucs-sep/ rather than in the shared Hugging Face cache, so a download stays scoped to the model you asked for.
This one is opt-in: switch it on in the Models page's “Enable more models” section before it appears in the picker.

Worth knowing

Requires exactly 44.1kHz. Anything else fails with a sample-rate mismatch, so Soundbox resamples the source first. The common case is splitting a 48kHz ACE-Step render.
The source goes in as the audio input, not the reference input. The reference input is for voice conversion and style transfer; using it here fails with an unrelated-sounding contract error.
Soundbox reads stem names and count from the output directory, because this has not been run end-to-end here. htdemucs emits vocals, drums, bass and other.

Sources

Primary sources: the model's own repository, the project it came from, and the conversion audio.cpp actually downloads. Never invented — an entry with nothing findable says so instead.

Licence details on these pages are summaries of the linked source documents, last checked 25 August 2026. They are not legal advice, and Soundbox is not affiliated with any model's authors — weights download from Hugging Face under the user's own account and terms. Company and model names identify whose work it is, and belong to their owners. The linked licence text is authoritative; commercial use should be checked against it, per model, before anything ships.