Bandit model (re)implementation as seen in "Remastering Divide and Remaster: A Cinematic Audio Source Separation Dataset with Multilingual Support" by Karn N. Watcharasupat, Chih-Wei Wu, and Iroro Orife. [arXiv].
Model weights are available on Zenodo here.
PS: I'm hoping to get this repo to eventually unify all Bandit-based models, but research code gets refactored incompatibly a lot so let's see where this goes.