BSRNN

Band-Split RNN is the paper every current separator cites. Its insight was musical rather than architectural: different instruments live in different frequency regions with different statistics, so instead of modeling 257 identical STFT bins, split the spectrum into hand-designed subbands — narrow where fundamentals and low harmonics live, wide where they do not — normalize each band separately, then alternate a band-axis RNN with a time-axis RNN to estimate complex masks.\n\nTrained on MUSDB18-HQ alone, the authors report it significantly beat the top-ranking models from the Music Demixing Challenge 2021, and a semi-supervised finetuning stage pushed all four stems further. The design then became the backbone of everything: BS-RoFormer and Bandit generalize it, and its speech-domain sibling powers multiple DNS Challenge systems.\n\nThere is no official weight release — the ecosystem runs on community reimplementations (PyTorch, SpeechBrain ports). Treat the entry as the canonical research design: peer-reviewed at IEEE/ACM TASLP in 2023, massively influential, but you build or borrow the weights yourself.

• Author-reported (arXiv 2209.15174, IEEE/ACM TASLP 2023): trained only on MUSDB18-HQ, significantly outperforms top-ranking MDX Challenge 2021 models; semi-supervised finetuning improves all four stems\n- Design basis for BS-RoFormer, Bandit, and multiple speech-enhancement challenge systems (community-validated lineage)\n- No official weights released — no directly usable benchmark artifact