While novelty is on the weak side for NeurIPS, reviewers found it overall simple and elegant. Results on synthetic data are found impressive, and proposed audio clips seem promising. Authors addressed most of the reviewers concerns with their rebuttal. The only major point remaining is how one could benchmark on real-world data. Two possible paths were suggested by reviewers during the discussion: i) use the DNS dataset (but this one was published only 4 months ago) ii) show improvement on a downstream task like ASR - but this would require quite a bit of work and could be viewed as been out-of-scope for this paper. The consensus was thus overall leaning towards acceptance.