Split-Attention Mechanisms with Graph Convolutional Network for Multi-Channel Speech Separation

In this paper, we introduce a split-attention mechanism integrated with a graph convolutional network for the task of multi-channel speech separation. Our approach involves dividing the embedding space of an encoder-decoder architecture into distinct branches. Within each branch, we employ a graph convolutional network (GCN) to capture spatial dependencies among various channels (nodes). This framework enables us to harness the synergistic benefits of both feature-map attention and multi-path representation by channel-wise attention across diverse network branches. To validate our method, we conducted extensive experiments on the spatialized WSJ0-2MIX dataset. The experimental results clearly demonstrate the superior performance of our approach compared to the baseline Beam-Guided TasNet, indicating its potential for enhancing speech separation tasks.