A lightweight, questionably‑engineered wrapper around Demucs. Built for people who “just want to split a track” without reading 40 pages of academic papers or compiling PyTorch at 3 AM.
Because Demucs is amazing, but:
- sometimes you just want to throw a file at it and hope for the best
- you don’t want to memorize the CLI every time
- you don’t want to pretend you understand complicated model names
- and mostly because no one had made a tool that openly admits it probably works
- Accepts an audio file
- Sends it to Demucs for separation
- Crosses fingers
- Returns stems (vocals, bass, drums, etc.)
- Improve Demucs
- Be faster than Demucs
- Be smarter than Demucs
- Replace Demucs
If the output sounds odd, that’s likely Demucs — not this wrapper.
- Python (for the Demucs worker)
- Docker & docker-compose (recommended for easy setup)
- A bit of patience (Demucs can be slow on CPU)
Clone the repo, then use docker-compose:
git clone <repo-url>
cd project-demucs
docker-compose build
docker-compose up -dNote: building the Demucs image can take time and requires network access to download Python packages and models.
Open the web UI at: http://localhost:3000
Upload a file and choose a model. The backend will save the file to ./uploads and the Demucs worker will process it, writing stems to ./output.
backend/ — Node.js backend (serves UI, handles uploads)
demucs/ — Demucs worker image (watches uploads/ and writes to output/)
uploads/ — incoming files (mounted into containers)
output/ — generated stems and status JSON files
The worker writes per-file status JSON into output/status/.json. The web UI polls this file to show progress and logs.
If you find this useful, you can support the project:
Pull requests welcome. Especially if they:
add more helpful messages (or sarcasm), remove unnecessary code, or make the project look more serious than it actually is.
MIT