Demucs
More info
- Creator
- Alexandre Défossez (Meta AI)
- Released
- Type
- Library
- License
- MIT
- Source
- github.com/facebookresearch/demucs
Meta's open model for splitting a finished mix into vocals, drums, bass and everything else — now the de facto engine behind much of the free stem separation out there.
Alexandre Défossez introduced Demucs at Facebook AI Research in 2019. What set it apart was that it worked directly on the waveform, with a U-Net-style architecture, when nearly everyone else was crunching a spectrogram. Later versions did both: Hybrid Demucs in 2021 and Hybrid Transformer Demucs in 2022, where a transformer ties the time and frequency branches together.
The code is PyTorch and MIT-licensed, which explains why it turns up embedded in DAWs, web services and karaoke apps that rarely mention where it came from. Development at Meta has wound down, but the models remain the benchmark others are measured against.