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

Researchers propose MeCo, a MeanFlow-based generative corrector that improves multi-channel speech separation by refining discriminative model outputs in a single step. The method combines Data-Space Optimization with specialized loss functions to achieve state-of-the-art performance in both signal fidelity and human listening quality with minimal computational cost.

MeCo addresses a fundamental challenge in speech separation technology: the gap between metrics that machines optimize for and what humans actually perceive as good audio quality. Traditional discriminative models excel at reference-based metrics like SI-SDR but often produce speech that sounds unnatural or degraded to listeners. This new approach introduces a generative correction layer that acts as a refinement stage, mapping imperfect machine estimates onto a manifold of clean speech characteristics.

The technical innovation centers on learning a conditional velocity field that guides estimates toward natural speech in just one generation step. By combining an $\mathbf{x}_r$-loss that penalizes errors across longer displacement intervals with an Endpoint SI-SDR loss focused on signal fidelity, the researchers create a dual-objective function that balances perceptual quality with acoustic accuracy. This represents a meaningful evolution in how researchers think about the separation problem—not as a single-stage task but as a two-stage refinement where initial estimates get polished for human consumption.

For the audio processing and speech technology industries, this work has practical implications. The minimal computational overhead means the correction step could be integrated into existing systems without significant infrastructure changes. The consistent improvement across both in-domain and out-of-domain scenarios suggests the approach generalizes well, potentially enabling deployment in real-world applications where training and deployment conditions differ.

As speech separation increasingly powers accessibility tools, hearing aids, and communication platforms, the shift toward optimizing for human perception rather than metric optimization represents an important maturation of the field. Future work will likely explore whether similar correction frameworks apply to other audio processing tasks.

• →MeCo uses a one-step generative corrector to refine speech separation estimates, bridging the gap between machine metrics and human listening quality
• →Data-Space Optimization combines displacement-based and signal-fidelity losses to simultaneously optimize perceptual and acoustic quality
• →The method achieves state-of-the-art results with minimal computational overhead, enabling practical deployment
• →Performance improvements hold consistently in both controlled and out-of-domain real-world scenarios
• →The approach signals an industry shift from metric-driven optimization toward perceptually-aligned audio processing