Music Genre Classification using Residual Attention Network

This paper proposes a new method in music genre classification by using the Residual Attention Network (RAN). Through the integration of attention mechanism and stacking attention modules, RAN has shown efficiency in the field of image processing. Residual blocks and attention modules are the most important considerations in RAN. First, each audio file is converted into a set of spectral images. Next, RAN is used for the classification of these spectral images, where the optimal stochastic gradient descent algorithm is used as the training model. The method is evaluated on the music data set from the Zalo AI “Music Genre Classification” challenge. This data set includes recordings of 10 different musical genres in Vietnam. Data augmentation combined with error analysis were implemented on the validation set. The results obtained had an accuracy of 71.7% on the test set. This highlights the potential of the proposed method in audio recognition applications.