Speech denoising models typically need clean reference audio for supervised training, but synthetic clean-noisy pairs can fail to match real-world conditions. Unsupervised approaches like MixIT and noisy-target training sidestep this by learning directly from noisy recordings, yet they remain noticeably less accurate than supervised systems.
The arXiv paper introduces SNOAR, a self-training framework that adds noise filtering, observation adding, and remixing to the unsupervised pipeline. The abstract describes this as a way to improve denoising without requiring clean references, directly addressing the sim-to-real gap while pushing unsupervised performance closer to supervised levels.
Because this is a single source, there are no conflicting findings to compare. The abstract does not report specific numbers or experimental details, so the significance here is the proposed mechanism itself: a self-training recipe that may make unsupervised denoising more practical for real-world audio.