MeCo: One-Step MeanFlow-based Corrector for Multi-Channel Speech Separation
2026-06-08 • Artificial Intelligence
Artificial Intelligence
AI summaryⓘ
The authors designed a new method called MeCo to improve how computers separate different voices in recordings, focusing on making the output sound better to human ears, not just score better on tests. MeCo works by learning a way to clean up noisy speech in a single step, guided by something called a MeanFlow. They also created a training approach named Data-Space Optimization that helps the system focus on both the overall quality and the exact final sound. Their experiments show that MeCo gives better results than previous methods without needing much extra computing power. This works well even for recordings the system wasn't specially trained on.
multi-channel speech separationdiscriminative modelsMeanFlowgenerative correctorData-Space Optimizationx_r-lossSI-SDR lossclean speech manifoldsignal fidelityhuman listening quality
Authors
Dohwan Kim, Jung-Woo Choi
Abstract
While discriminative models for multi-channel speech separation excel in reference-based metrics, they often exhibit suboptimal human listening quality. To address this, we propose a novel MeanFlow-based one-step generative corrector (MeCo). MeCo learns a conditional average velocity field to map discriminative estimates directly onto the clean speech manifold in a single step. To maximize one-step generation performance, we introduce Data-Space Optimization (DSO). DSO integrates an $\mathbf{x}_r$-loss, which penalizes prediction errors on longer displacement intervals to serve as a generative objective for human listening quality, with an Endpoint SI-SDR loss that directly optimizes terminal signal fidelity. Experiments demonstrate that MeCo achieves state-of-the-art (SOTA) performance with minimal computational overhead, simultaneously achieving superior signal fidelity and human listening quality in both in-domain and out-of-domain scenarios.