Abstract: Abstract: In the field of speech processing， multi-channel speech separation technology aims to effectively separate the speech signals of different speakers from multi-channel mixed speech. However， existing methods have shortcomings in handling long-distance dependency relationships between multi-channel feature points. In response to this issue， this study proposes a novel multi-channel speech separation method based on self-guided Transformer （SG-former）， which aims to construct an adaptive fine-grained global attention mechanism. The core mechanism of SG-former is to reassign tokens through saliency maps. Under this framework， significant regions can extract key information at a fine-grained level， while secondary regions adopt a coarse-grained extraction approach to reduce computational costs. The generation of saliency maps relies on the hybrid-scale self-attention mechanism， which can accurately capture the long-distance dependency relationships between multi-channel feature points. To verify the effectiveness of the proposed method， experiments were conducted on a spatialized WSJ0-2MIX database. The experimental results show that the SG-former method has a significant advantage in Signal to Distortion Ratio Improvement （SDRi） compared to the baseline Beam Guided TasNet method， achieving a 20.34 dB improvement. This result fully demonstrates the superiority of SG-former in dealing with multi-channel speech separation problems， especially in establishing long-distance dependency relationships. The experimental results show that this method outperforms existing technologies in terms of performance， providing new ideas and methods for research in the field of multi-channel speech separation.

Key words: Key words: speech separation, self-guided attention, hybrid-scale attention, multi-channel processing, beamformer