ADS-BiMamba: Attentive Dynamic-Split Bidirectional Mamba for Multi-Channel Speech Enhancement

In multi-channel speech enhancement, Mamba is attractive because of its strong modeling capability and linear computational complexity. However, recurrent information propagation restricts direct interactions between non-adjacent temporal or spectral elements, limiting the extraction of global context even with bidirectional scanning. We propose ADS-BiMamba, a bidirectional Mamba with an attentive dynamic split mechanism. In the backward branch, a split predictor estimates a split point and reconfigures the sequential order to change scan adjacency and shorten the recurrent propagation path for non-contiguous temporal or spectral positions. We design an information fusion module to fuse forward and backward features, aggregate long-range information, and preserve linear computational complexity. Experiments show that ADS-BiMamba achieves 3.67 PESQ on spatial DNS dataset and a Metric of 0.982 on the L3DAS22 dataset with 3.9 M parameters and 41.7 G MACs, improving the clarity and intelligibility of enhanced speech.