Parallel attention of representation global time–frequency correlation for music genre classification

Music genre classification (MGC) is an indispensable branch of music information retrieval. With the prevalence of end-to-end learning, the research on MGC has made some breakthroughs. However, the limited receptive field of convolutional neural network (CNN) cannot capture a correlation between temporal frames of sounding at any moment and sound frequencies of all vibrations in the song. Meanwhile, time–frequency information of channels is not equally important. In order to deal with the above problems, we apply dual parallel attention (DPA) in CNN-5 to focus on global dependencies. First, we propose parallel channel attention (PCA) to build global time–frequency dependencies in the song and study the influence of different weighting methods for PCA. Next, we design dual parallel attention, which focuses on global time–frequency dependencies in the song and adaptively calibrates contribution of different channels to feature map. Then, we analyzed the effect of applying different numbers and positions of DPA in CNN-5 for performance and compared DPA with multiple attention mechanisms. The results on GTZAN dataset demonstrated that the proposed method achieves a classification accuracy of 91.4%, and DPA has the highest performance.