Demucs

Demucs is Meta AI Research (FAIR)'s music source separation model — feed it any song and it splits it into drums, bass, vocals, and "other" stems. The current v4 is Hybrid Transformer Demucs: a hybrid spectrogram/waveform U-Net whose innermost layers were replaced with a cross-domain Transformer encoder (self-attention within each domain, cross-attention across domains).

Benchmarks (provider-reported, repo README + paper arxiv:2211.08553):

• 9.00 dB SDR on the MUSDB HQ test set
• 9.20 dB SDR with sparse-attention kernels (extended receptive field) and per-source fine-tuning — claimed state-of-the-art at release
• (SDR = signal-to-distortion ratio, higher is better; the repo's comparison table shows it ahead of Wave-U-Net, Open-Unmix, D3Net, and Conv-TasNet on overall SDR)

Practical notes: it's a CLI/Python tool (PyPI: demucs), not an API — you run it locally on audio files. With 10.3k GitHub stars it's the backbone of most open vocal-remover/karaoke/remix pipelines. Legal caveat: weights are MIT, but the model was trained on MUSDB HQ, whose dataset license is CC BY-NC-SA — worth a legal review for commercial use.

Provider-reported (repo README + HTDemucs paper arxiv:2211.08553):

• 9.00 dB SDR on MUSDB HQ test set (v4 default)
• 9.20 dB SDR with sparse-attention kernels + per-source fine-tuning

No independent re-verification run by me; the paper is peer-published and widely cited.