Demucs / HTDemucs

About

Meta FAIR's music source separation model family. Waveform-domain U-Net with bidirectional LSTM (2019), evolved through Hybrid Demucs (2021) to Hybrid Transformer Demucs (2022). State-of-the-art at publication (9.20 dB SDR on MUSDB18HQ). Provides pretrained 4-stem and 6-stem models. GitHub: facebookresearch/demucs (10.1k stars). Archived by Meta Jan 2025 but still fully functional.

Relevance

Best general-purpose separation tool for bluegrass recordings. Separates into drums/bass/vocals/other — banjo and other string instruments land in "other." Installable via pip install demucs.

Mentions

• ../sources/2026-05-19-demucs-htdemucs — ingested summary
• ../entities/bs-roformer — current SOTA that outperforms HTDemucs
• ../entities/spleeter — faster but lower quality alternative
• ../entities/open-unmix — more hackable/trainable alternative