stemsplit

Two small command-line tools for producers:

• stemsplit splits a song into stems (vocals, drums, bass, other) with Demucs. Works with a local file or any URL yt-dlp supports.
• ytsmpl downloads a track as WAV without splitting it, for use as a sample.

Both tools detect the track's tempo (BPM) and musical key, add them to the file names (vocals_121bpm_Am.wav), and write them as ID3 tags (TBPM / TKEY) so DAWs and sample browsers can read them.

Requirements

• Linux or macOS
• Python 3.10 to 3.13
• ffmpeg
• aubio command-line tools (provides aubiotrack, used for BPM detection)
• yt-dlp, a recent release (2025 or later; ytsmpl uses its -t wav preset)

Install the system tools:

# Debian / Ubuntu
sudo apt install ffmpeg aubio-tools python3-venv
pipx install yt-dlp        # the apt package is usually too old

# Fedora
sudo dnf install ffmpeg-free aubio yt-dlp   # or ffmpeg from RPM Fusion

# Arch Linux
sudo pacman -S ffmpeg aubio yt-dlp

# macOS (Homebrew)
brew install ffmpeg aubio yt-dlp python@3.13

If your system Python is newer than 3.13, install a supported version with uv (uv python install 3.13) or pyenv.

Installation

git clone https://github.com/CliveBanner/stemsplit.git
cd stemsplit

# The virtual environment must be called "venv" and live in this folder:
# the scripts look for it there and switch to it automatically.
python3 -m venv venv     # or e.g. python3.13 -m venv venv
venv/bin/pip install -r requirements.txt

# Put both commands on your PATH
mkdir -p ~/.local/bin
ln -s "$PWD/stemsplit.py" ~/.local/bin/stemsplit
ln -s "$PWD/ytsmpl.py"    ~/.local/bin/ytsmpl

Make sure ~/.local/bin is on your PATH. Check it works:

stemsplit --help

The first split downloads the Demucs model (about 80 MB) to ~/.cache/torch.

GPU and CPU-only builds

pip picks the PyTorch build for your platform. On Linux that is usually a CUDA build, which uses an NVIDIA GPU automatically if one is available. On a Mac, Demucs runs on the CPU. Without a GPU, splitting still works but is slow (roughly the length of the song, or longer).

If you have no NVIDIA GPU and want a much smaller download, install the CPU-only build first:

venv/bin/pip install "torch<2.9" "torchaudio<2.9" --index-url https://download.pytorch.org/whl/cpu
venv/bin/pip install -r requirements.txt

For an older NVIDIA GPU, pick the PyTorch build that matches your driver's CUDA version from pytorch.org. Check that PyTorch sees your GPU:

venv/bin/python -c "import torch; print(torch.cuda.is_available())"

Usage

stemsplit

stemsplit https://www.youtube.com/watch?v=...     # 4 stems from a URL
stemsplit /path/to/song.mp3                       # 4 stems from a local file
stemsplit -s 2 /path/to/song.wav                  # vocals + instrumental only
stemsplit --mp3 -o ~/stems /path/to/song.flac     # 320k MP3 to a custom folder

Option	Description	Default
-s, --stems	2 (vocals / no_vocals) or 4 (vocals / drums / bass / other)	4
-o, --output	Output folder	~/Music/Stems
--mp3	Write 320k MP3 instead of WAV	off

Output goes to <output>/htdemucs/<song name>/:

htdemucs/My Song/
├── bass_121bpm_Am.wav
├── drums_121bpm_Am.wav
├── other_121bpm_Am.wav
└── vocals_121bpm_Am.wav

Running the same song again overwrites its stems.

ytsmpl

ytsmpl https://www.youtube.com/watch?v=...
ytsmpl -o ~/samples https://www.youtube.com/watch?v=...

Option	Description	Default
-o, --output	Output folder	~/Music/Samples

The WAV is saved as <title>_<bpm>bpm_<key>.wav, with the same tags.

How detection works

• BPM: aubiotrack finds beat positions; the tempo comes from the median gap between beats. Half- or double-time results are possible on some tracks.
• Key: librosa computes the average chroma of the track, which is compared against the Krumhansl-Schmuckler major and minor key profiles. Keys are written as C, F#, Am and so on. Treat the result as a good first guess.

If detection fails, the files are saved without the BPM/key label.

Troubleshooting

• aubiotrack: command not found: install the aubio tools (see Requirements).
• yt-dlp errors: update it (pipx upgrade yt-dlp or your package manager). Sites change often, and old versions stop working.
• No module named ...: the venv folder is missing or in the wrong place. It must be inside the cloned folder and named exactly venv.

Legal

Only download and process audio you have the rights to use. Respect the terms of service of the sites you download from.

License

Public domain (Unlicense), see LICENSE.