Bandit

Bandit generalizes the Band-Split RNN idea to arbitrary and overcomplete frequency partitions, using psychoacoustically motivated band definitions (Bark, ERB, mel) with redundancy for more reliable feature extraction. Its target is cinematic audio source separation: splitting a soundtrack into dialogue, music, and effects stems. A shared encoder feeds detachable per-stem decoders, which cuts compute during both training and inference and lets you drop in only the decoders you need.\n\nThe authors report state-of-the-art results on the Divide and Remaster (DnR) dataset, claiming performance above the ideal ratio mask on the dialogue stem — an unusually strong result. For music specifically, the same lab's query-based sibling model, Banquet, does language-queried music stem separation.\n\nThe code is Apache 2.0, but note the weights: the official v1 checkpoints sit on Zenodo under CC-BY-NC-4.0, so commercial use of the pretrained weights needs a different arrangement. It is a research model with honest scope — cinematic stems, not the classic vocals/drums/bass/other split — and it is best understood as one.

• Author-reported (IEEE Open J. Signal Process. 2024, doi 10.1109/OJSP.2023.3339428): state-of-the-art on the Divide and Remaster dataset; above ideal-ratio-mask performance on the dialogue stem (provider-claimed)\n- Author-reported complexity: ~33-83M params depending on band config (repo benchmark table)\n- Cinematic task (dialogue/music/effects), not the MUSDB18 4-stem task — numbers are not comparable to music separators