Demucs
A command-line model that splits a song into stems — typically drums, bass, vocals, and the rest of the accompaniment.
- Type Stems
- Difficulty Medium
- Hardware Python 3.10 or newer (some current setups need Python 3.12 at the highest). A GPU is recommended: at least 3 GB of GPU RAM for Demucs, and about 7 GB with the default settings. You can lower memory use with a smaller segment size, or run on CPU. On CPU, the README says processing time is roughly 1.5 times the length of the track.
- License MIT (code); weights license unclear
What it does
Demucs is a music source-separation model originally developed at Facebook / Meta AI by Alexandre Défossez and collaborators. The original Meta repository is archived. The official repo to use now is Défossez's own fork. He also writes that he is not actively working on Demucs, so replies may be slow and new features should not be expected.
The current release is Hybrid Transformer Demucs. In normal use it separates drums, bass, vocals, and other. An experimental six-source model adds guitar and piano; the README says guitar is okay in quick tests, and piano has bleeding and artifacts.
If you want a desktop interface instead of the terminal, Ultimate Vocal Remover can run Demucs models.
Why it matters for hip-hop
A clean drum stem, an isolated bass, or an acapella is everyday raw material: chops, remixes, practice mixes, and checking your own bounces. Demucs is one of the better-known open models for that job, and it runs locally.
Splitting a commercially released song does not grant you the right to use those stems on a record you sell.
Getting it running
The usual path is a Python environment, installing the Demucs package, and running it on a track. That is Medium: not hard if you have used Python, but not a one-click installer. If package and GPU drivers are a hassle, use Ultimate Vocal Remover and pick a Demucs model there.
License notes
This one needs a careful reading. The code is MIT. The README is blunt about the weights: the license of the weights is not clear, because training involved dataset material that was limited to research use, plus Meta internal data. The author writes that he cannot grant a license for those weights, and that one six-source variant likely falls under a non-commercial Creative Commons-style term. The program code is permissively licensed; the pretrained models are not clearly cleared for every commercial use. If you need certainty for a commercial product or a company workflow, do not assume the weights are MIT just because the code is. That is the project's own warning.