spleeter

Spleeter is an open-source music source separation library developed by Deezer Research. Written in Python and powered by TensorFlow, it provides pretrained state-of-the-art models that can decompose a mixed audio track into individual stems. Supported separation modes include 2-stem (vocals and accompaniment), 4-stem (vocals, drums, bass, other), and 5-stem (vocals, drums, bass, piano, other). The 2-stem and 4-stem models achieve high performance on the MUSDB dataset, and on GPU hardware Spleeter can process audio up to 100x faster than real-time. The library is designed for both command-line usage and integration into custom Python pipelines. Installation is available via pip or Docker, and a Google Colab notebook is provided for quick experimentation without local setup. Dependencies include ffmpeg and libsndfile, installable through Conda. The project is managed with Poetry for development and testing. Spleeter's pretrained models have been adopted by professional audio software including iZotope RX 8 (Music Rebalance), Steinberg SpectralLayers 7 (Unmix), Acon Digital Acoustica 7, VirtualDJ stem isolation, and Algoriddim's NeuralMix and djayPRO. It also serves as a baseline in the Music Demixing Challenge at ISMIR 2021. A commercial version, Spleeter Pro, is available through Deezer Tech Services for users needing faster processing and dedicated support. The project is MIT-licensed and published in the Journal of Open Source Software (JOSS). Known limitations include compatibility issues with Apple M1 chips due to TensorFlow constraints, and occasional CLI shortcut problems on Windows that can be worked around by invoking the module directly via python -m spleeter.