Columbia researchers fix AI speech suppression with inference-time fix

A rethink-and-refine correction module tackles over-suppression in speech enhancement models, where speech gets removed alongside noise. It works entirely at inference time, needing no extra training, so it can plug into diverse enhancement systems. Using an ASR model to get word- or phoneme-level alignments, it spots unreliable intervals and selectively remixes them via convex interpolation, with per-segment weights tuned to balance perceptual quality and speech preservation. Tests on the URGENT 2024/2025, VCTK-DEMAND, and MSP-PODCAST datasets show consistent gains in perceptual quality, intelligibility, and downstream performance over standard enhancement alone, highlighting the framework's value for robust speech processing.