SPATIAL-TEMPORAL GRAPH CONVOLUTION NETWORK FOR MULTICHANNEL SPEECH ENHANCEMENT

Spatial dependency related to distributed microphone positions is essential for multichannel speech enhancement task. It is still challenging due to lack of accurate array positions and complex spatial-temporal relations of multichannel noisy signals This paper proposes a spatial-temporal graph convolutional network composed of cascaded spatial-temporal (ST) modules with channel fusion. Without any prior information of array and acoustic scene, a graph convolution block is designed with learnable adjacency matrix to capture the spatial dependency of pairwise channels. Then, it is embedded with time-frequency convolution block as the ST module to fuse the multi-dimensional correlation features for target speech estimation. Furthermore, a novel weighted loss function based on speech intelligibility index (SII) is proposed to assign more attention for the important bands of human understanding during network training. Our framework is demonstrated to achieve over 11% performance improvement on PESQ and intelligibility against prior state-of-the-art approaches in multi-scene speech enhancement experiments.