Deezer's open-source stem splitter that powers pro DAWs

What it does

Spleeter is a Python library that splits mixed audio into isolated stems using pretrained TensorFlow models. It ships with ready-made models for separating tracks into two stems (vocals and accompaniment), four stems (vocals, drums, bass, and other), or five stems (adding piano). You can embed it in a Python pipeline or invoke it directly, and it can separate audio into four stems roughly 100× faster than real time on a GPU.

The interesting bit

The pretrained models have been absorbed into commercial tools you probably know: iZotope RX 8, Steinberg SpectralLayers 7, VirtualDJ, and Algoriddim’s djayPRO suite all use Spleeter under the hood. There is even a community port for Ableton Live. For an open-source research project, that is a lot of time in professional DAWs.

Key highlights

• Ships with pretrained 2-stem, 4-stem, and 5-stem separation models (vocals, drums, bass, piano, and other)
• Can be dropped into a Python pipeline or invoked as a standalone tool
• Runs 4-stem separation roughly 100× faster than real time on a GPU
• Serves as a baseline model for the Music Demixing Challenge
• MIT-licensed, with a commercial Spleeter Pro tier available separately

Caveats

• Version 2.1.0 introduced breaking CLI changes and dropped the dedicated GPU package, so existing scripts may need updates
• Apple M1 chips have known TensorFlow compatibility issues, and the maintainers no longer recommend installing via Conda
• The Windows packaged shortcut is currently unreliable

Verdict

Worth a look if you are building audio pipelines or research prototypes that need fast stem separation. Look elsewhere if you want a polished native GUI or flawless Apple Silicon support today.