Urdu Music Genre Classification Using Convolution Neural Networks

Abstract

Music is a form of art whose medium is organized sound. Communities from all parts of the world can be identified by their songs they listen to and compose. Music is often classified into different genres. The goal of this paper is to discover a machine learning algorithm that efficiently classifies music. Our research focuses on the genre classification of Urdu music, which is native to Pakistan. In order to compare performance, different classification models were constructed and trained over the Urdu dataset we created using several methods, mainly focusing on convolution neural networks. This dataset was based on only Mel-Frequency Cepstral Coefficients images of audio tracks. The performances were compared in terms of validation accuracies of models and loss produced by Sparse Categorical Cross Entropy Loss function. Our model based on CNN with batch normalization gave us a higher level of accuracy of 92.6% with a low loss of 0.0051, compared to other methods.