Music genre classification: Machine Learning on GTZAN

Abstract

This paper explores music genre classification, aiming to enhance existing methodologies. As a crucial aspect of music information retrieval, genre classification facilitates organization and recommendation in music databases and streaming services. Our research, inspired by a Kaggle project, examines the background of music genre classification and introduces improvements. The study focuses on data preparation techniques and a novel methodology using Support Vector Machines (SVM). Utilizing the GTZAN dataset, we applied data segmentation and feature extraction, employing machine learning algorithms like Logistic Regression, Random Forest, and SVM. A significant innovation is our segmentation technique based on music's beats per minute (BPM), designed to preserve rhythmic structure, believed to be essential for accurate classification. We explored various feature extraction methods to boost classifier performance. Experimental results showed the 3-second segmented dataset performed better with SVM's linear kernel. Additionally, a 4-beat segmentation experiment suggested that finer segmentation captures richer audio features, potentially improving classification accuracy. The paper concludes with findings and future research directions, including dataset expansion, advanced segmentation based on musical theory, deep learning applications, and developing real-time classification systems.