MeCo: One-Step MeanFlow-based Corrector for Multi-Channel Speech Separation

The paper introduces MeCo, a one-step MeanFlow-based generative corrector for multi-channel speech separation, which enhances human listening quality compared to traditional discriminative models. It employs Data-Space Optimization (DSO) with an $\mathbf{x}_r$-loss and Endpoint SI-SDR loss to improve both generative performance and signal fidelity. MeCo achieves state-of-the-art performance with minimal computational overhead, making it a valuable tool for practitioners focused on improving the quality of speech separation systems.