Get Started
← Creator toolkit · Mission 8 of 13

Remove the vocals from any song in 2 minutes, free

Demucs, the free AI model first built at Meta's research lab, pulls a song apart into vocals, drums, bass and everything else. Karaoke tracks, a cappellas, practice tracks and a finished karaoke video, all on your own laptop with no uploads and no credits.

35 minBeginnerTools: FreeStep 1 free
Tested 9 Oct 2026 · Demucs 4.1.0 (htdemucs model), PyTorch 2.14.1 on Python 3.12 via uv 0.11.9, 16GB laptop with no graphics card, Windows 11
Demucs installeda free, open-source AI that splits music into separate parts (called stems)
Vocals, drums, bass and everything else as four separate files, from one command
A karaoke version and an a cappella of any song, as MP3s
  • A vertical karaoke video with the lyrics on screen, ready for Reels or TikTok
  • About 2 minutes for a 3-minute song on a laptop with no graphics card, and a clear picture of where it leaks
Before you start
  • Any recent laptop. Demucs uses at most about 1.4GB of RAM
  • About 1GB of free disk space (the tool, one 84MB model and room for the files it makes)
  • FFmpeg for the final video (the FFmpeg lesson installs it)
  • uv installed. It's a small, official tool that installs Python apps cleanly. Step 1 shows you how
  • A song you're allowed to use, or the free Creative Commons track linked in step 2
01/06Install Demucs0 done
STEP 01 of 06~5 min

Install Demucs

One command installs Demucs in its own sealed box, so it can't break anything else on your PC.

  1. Demucs was created by Alexandre Défossez and colleagues at Meta's research lab. The original Facebook Research copy on GitHub is now archived (read-only). Défossez has since left Meta and keeps it going at github.com/adefossez/demucs ↗, which calls itself the officially maintained Demucs. It's MIT-licensed, and it's what you get when you install demucs today.

  2. You install it with uv. If you don't have uv yet, install it first, then open a new terminal:

    Windows (only if you don't have uv)
    winget install --id astral-sh.uv -e
    Mac (only if you don't have uv)
    brew install uv
    Mac or Linux without Homebrew (only if you don't have uv)
    curl -LsSf https://astral.sh/uv/install.sh | sh

  3. Now install Demucs. Copy the whole line. The --with numpy part fixes a missing piece in Demucs 4.1.0 (see the warning below), and --python 3.12 pins a Python version that works with it. uv downloads it for you if needed:

    Terminal
    uv tool install demucs --with numpy --python 3.12
    Output
    ...
 + sphn==0.2.1
 + sympy==1.14.0
 + torch==2.14.1
 + tqdm==4.70.1
 + truststore==0.10.4
 + typing-extensions==4.16.0
Installed 1 executable: demucs

  4. The first install takes a minute or two and uses about 595MB of disk. Most of that is PyTorch, the engine that runs the AI.

    CarefulThe official one-liner (uv tool install demucs, or uvx demucs) installs without complaint, then crashes the first time you use it with ModuleNotFoundError: No module named 'numpy'. Version 4.1.0 forgot to list numpy as something it needs. A fix is waiting on GitHub, but until it ships, --with numpy is the cure.

  5. Check it worked:

    Terminal
    demucs --help
    Output
    usage: demucs.separate [-h] [-s SIG | -n NAME] [--repo REPO] [--list-models]
                       [-v] [-o OUT] [--filename FILENAME] [-d DEVICE]
...
✓
Checkpointdemucs --help prints a list of options starting with usage: demucs.separate.
If it breaksModuleNotFoundError: No module named 'numpy' + 2 more
ModuleNotFoundError: No module named 'numpy'
You installed it without the fix. Run uv tool uninstall demucs, then the install command above (with --with numpy).
'demucs' is not recognised
uv puts its apps in a folder your terminal doesn't know about yet. Run uv tool update-shell, close the terminal, open a new one and try again.
You have an NVIDIA graphics card and want more speed
Everything here runs on the CPU, so it works on any laptop. The project's Windows guide ↗ shows how to install the graphics-card version of PyTorch instead.
Still stuck? Ask the community
Next mission · Creator toolkitCut the silences out of any talking video automatically
→