Demucs

Demucs, developed at Meta research, separates a mixed music track into individual stems: drums, bass, vocals, and other accompaniment, with an experimental six-source model adding guitar and piano. The fourth version, Hybrid Transformer Demucs, processes spectrogram and waveform representations in parallel through a dual U-Net joined by cross-domain transformer layers, reaching 9.0 dB SDR on the MUSDB HQ benchmark and 9.2 dB with fine-tuning. Separation runs on a GPU with as little as 3 GB of memory, 7 GB recommended, or on CPU at roughly 1.5 times the audio duration. The original repository is no longer maintained after the author left Meta, and continued bug-fix work happens in a community fork, but the MIT-licensed code and pretrained models remain widely used. Musicians, producers, audio engineers, and researchers rely on it for remixing, karaoke tracks, sampling, and dataset preparation.