Demucs / HTDemucs

About

Music source separation system from Meta FAIR (Facebook AI Research). Evolved across three major versions: Demucs (waveform U-Net + BiLSTM, 2019), Hybrid Demucs (spectrogram + waveform, 2021), and HTDemucs (Hybrid Transformer Demucs, 2022). All versions available in a single GitHub repository. As of 2025, the repo is archived by Meta but the code and pretrained models remain functional.

Relevance

Best general-purpose source separation tool available via pip install demucs. Provides 4-stem (vocals/bass/drums/other) and 6-stem (+guitar/piano) models. The "other" stem is a catch-all — insufficient for separating banjo from mandolin from fiddle in bluegrass. For bluegrass, the most practical use is as a first-pass separator, with the "other" stem fed to additional processing.

Mentions

• ../sources/2019-11-27-demucs — original Demucs paper
• ../sources/2022-11-15-htdemucs — HTDemucs paper
• ../entities/bs-roformer — current SOTA, surpassed HTDemucs
• ../entities/open-unmix — more hackable alternative

Links

• GitHub: https://github.com/facebookresearch/demucs (10.1k stars, archived)
• Install: pip install demucs