SCNet

SCNet attacks the efficiency problem in wide-band music separation. The observation: subbands carry very different amounts of information, so compressing them all equally wastes capacity. SCNet splits the spectrogram into subbands and applies a sparsity-based encoder that compresses information-sparse bands harder, densifying the signal where it matters and spending modeling budget on the bands that carry the music.\n\nThe authors report 9.0 dB SDR on MUSDB18-HQ without any extra training data — a state-of-the-art claim at publication — with CPU inference at only 48% of HT Demucs' time. The large variant (SCNet-L) scores 10.86 vocals / 10.98 drums / 9.49 bass / 7.44 other in a third-party comparison table from the Moises-Light paper. ZFTurbo's Music-Source-Separation-Training framework ships trained SCNet checkpoints, which is how most people actually run it.\n\nThe paper was accepted at ICASSP 2024, the code is MIT, and checkpoints are public. It is a young line — the community is small and every headline number is still author-reported — but as an efficiency play it is one of the more interesting separators of 2024.

• Author-reported (arXiv 2401.13276, ICASSP 2024): 9.0 dB SDR on MUSDB18-HQ without extra data; CPU inference 48% of HT Demucs\n- Third-party (Moises-Light paper, arXiv 2510.06785): SCNet-L averages 9.69 SDR (vocals 10.86, drums 10.98, bass 9.49, other 7.44) on MUSDB18-HQ\n- Young model line; no independent replication published yet