1 January 2025 in IEEE Transactions on Audio, Speech and Language Processing
27 March 2026 in IEEE Signal Processing Letters
Generative models have recently shown strong potential for high-fidelity speech enhancement but suffer from high inference latency due to iterative sampling. In contrast, discriminative models offer fast processing but tend to introduce non-linear distortions that degrade downstream task performance. To bridge this gap, we propose MCFLOW-SE, a spacing-aware generative framework enabling one-step multichannel speech enhancement. By learning an average velocity field that directly maps noisy multichannel observations to clean speech, MCFLOW-SE enables fast one-step inference while effectively leveraging spatial information across channels. Furthermore, sensor spacing is incorporated as positional encoding into the generative process, improving robustness across the evaluated array-spacing configurations. Experiments on both simulated and real-world datasets demonstrate that MCFLOW-SE delivers substantially faster inference than prior generative approaches, while achieving superior Automatic Speech Recognition (ASR) accuracy over discriminative baselines. These results support its potential to bridge the gap between high-fidelity generation and real-time applicability.
1 January 2026 in IEEE Transactions on Instrumentation and Measurement