1.4k stars · MIT · 0.47.0 on PyPI, published 27 August 2026, year present. Code pushed 27 August 2026, proved via ungh.cc · Track this in Scout
One command that splits a recording into separate tracks, running the pre-trained separation models from the Ultimate Vocal Remover project.
▶Repo detailsthe review · specs · pros & cons · install
What it is
A command-line tool and a Python library that runs the pre-trained separation models from the Ultimate Vocal Remover project. It supports several model families, can combine two models for a better result, and handles long files by working through them in pieces.
What it is good for. Anyone who records over music, or who has a recording where two things are fighting. A podcast where the intro music is too loud under the first sentence. An interview with an air conditioner in the background. A video where you want the voice for subtitles and the music for nothing. A channel that records in two languages over one backing track can fix one of them without touching the other.
- MIT, free, and it downloads the models for you on the first run.
- It runs on a plain processor, and it also uses an NVIDIA graphics card or an Apple M-series chip if you have one.
- It is both a command and a library, so it fits into a script as easily as into your hands.
- The first run downloads a model, and the cache directory it uses by default is a temporary one, so a restart can mean downloading it again.
- It needs Python 3.10 or newer, and its own notes say Python 3.14.1 does not work.
- Separation is never perfect. A quiet voice under loud music comes back with artefacts, and no model fixes that.
- adefossez/demucs
One of the best-known separation models, used directly rather than through a wrapper, and it offers one family of models where this offers several.
Track this in Scout - deezer/spleeter
The tool from Deezer that made this possible for everyone. It is faster and older, and the newer models separate more cleanly.
Track this in Scout - Anjok07/ultimatevocalremovergui
The window-and-buttons version, from the project whose models this uses. Easier for one file, no good in a script.
Track this in Scout
python3 -m venv venv source venv/bin/activate pip install "audio-separator[cpu]" audio-separator /path/to/recording.wav
