A Comparative Performance Evaluation of Machine Learning Approaches for Spectrogram-based Music Genre Classification

The market for various music styles has expanded along with the company’s steadily expanding consumer base. It is crucial to categorize music according to genres in order to suit people’s needs. It is the responsibility of the listener to manually rank the music because it is a tedious and time-consuming operation. The music signals in this work are first transformed into the relevant spectrograms. The classifier then receives these spectrogram features as input. The work mainly uses a convolutional neural network. Different machine learning models are compared and validated against the neural network. The models will be trained and compared on GTZAN, which is a public dataset consisting of thousands of audio files comprising ten genres. The goal is to create a machine-learning model that categorizes music into its appropriate genre, assess the accuracy of this model against other classifiers, and derive the appropriate conclusions. The accuracies for K Nearest Neighbor, Multi-Layer Perceptron, Support Vector Machine, and Naive Bayes classifiers are 88.8%, 84.0%, 84.0%, and 51.8%, respectively. The accuracy that has been obtained using CNN is 94.87%, with a validation accuracy of 89.03% and a 0.84 f-score.