Multi-stage music separation network with dual-branch attention and hybrid convolution

In this paper, we propose a lightweight multi-stage network for monaural vocal and accompaniment separation. We design a dual-branch attention (DBA) module to obtain the correlation of each position pair and that among the channels of feature maps, respectively. The square CNN (i.e. the size of the filter is kx k) shares the weights of each of the square areas in feature maps that which makes its ability of feature extraction limited. In order to address it, we propose a hybrid convolution (HC) block based on hybrid convolutional mechanism instead of square CNN to capture the dependencies along with the time dimension and the frequency dimension respectively. The ablation experiments demonstrate that the DBA module and HC block can assist in improving the separation performance. Experimental results show that our proposed network achieves outstanding performance on the MIR-1K dataset only with fewer parameters, and competitive performance compared with state-of-the-arts on DSD100 and MUSDB18 datasets.