MeCo's one-step corrector boosts speech separation quality

Multi-channel speech separation models that rely on discriminative methods often score high on reference-based metrics but produce audio that sounds unnatural to human listeners. To bridge this gap, researchers Dohwan Kim and Jung-Woo Choi from KAIST have developed MeCo (MeanFlow-based Corrector), a one-step generative corrector that refines the output of any discriminative separator.

MeCo learns a conditional average velocity field that maps discriminative estimates directly onto the clean speech manifold in a single generative step. This contrasts with traditional diffusion models that require many steps. To maximize one-step generation performance, the team introduces Data-Space Optimization (DSO). DSO combines an x_r-loss—which penalizes prediction errors on longer displacement intervals to improve perceptual quality—with an Endpoint SI-SDR loss that directly optimizes the final signal's fidelity.

Experiments show MeCo achieves state-of-the-art results across multiple benchmarks, simultaneously improving both signal fidelity and human listening quality. The approach adds minimal computational overhead, making it practical for real-time applications. The paper has been accepted at Interspeech 2026 and is available on arXiv.