demucs

Demucs is Meta's music source separation model, working in the waveform domain to split a song into vocals, drums, bass and other stems — the research-grade sibling of the click-and-go separators.

The current generation is Hybrid Transformer Demucs (v4, htdemucs), which became the default model in the library and remains a strong baseline; the repo documents each architecture generation.

It ships as a pip package with a web demo on Hugging Face Spaces, and its results anchored the sound-demixing research community (the SDX challenge lineage).