Dynamic Spatial Filtering with Residual Spectral Mapping for Lightweight Multichannel Speech Enhancement

In this paper, we propose a novel multichannel speech enhancement method that is based on adaptive beamforming and spectral recovery. The core idea consists of dynamic spatial filtering (DSF) and residual spectral mapping (RSM). The DSF module utilizes depthwise separable convolutions to efficiently suppress directional noise and reverberation, while the residual spectral mapping (RSM) network refines the beam-formed output by restoring lost speech components. Experimental results on the CHiME 3/4 dataset demonstrate that the proposed model outperforms state-of-the-art methods in both speech quality and intelligibility while maintaining lower computational complexity.