NeurIPS 2020

Unsupervised Sound Separation Using Mixture Invariant Training

Meta Review

Reviewers agree that the proposed approach is simple, clever, and significantly novel. Experimental work shows it works well on synthetic data. The only major concern remaining after the rebuttal and discussion phases is the lack of evaluation on real-world data - while being able to leverage real data is claimed as an advantage of the approach. In the end, the consensus after the discussion period was the paper should be accepted, as the proposed idea is interesting enough.