Music Genre Classification Using Classical Machine Learning Algorithms on the GTZAN Dataset

Abstract

This study presents a comprehensive evaluation of classical machine learning algorithms for automatic music genre classification using the GTZAN dataset. We systematically analyze k-Nearest Neighbors (k-NN), Decision Trees, and Logistic Regression on 1,000 audio tracks spanning ten genres, utilizing 58 pre-extracted audio features including MFCCs, spectral descriptors, and temporal characteristics. Our methodology employs rigorous preprocessing, stratified cross-validation, and extensive hyperparameter optimization. Results demonstrate that k-NN achieves exceptional performance with 92.04% accuracy—significantly outperforming Logistic Regression (71.97%) and Decision Trees (65.12%). Per-class analysis reveals substantial genre-specific variations, with classical music achieving 94.97% accuracy while rock music presents the greatest challenge at 51.50%. Feature importance analysis identifies MFCC coefficients and spectral centroid as the most discriminative. Statistical significance testing confirms k-NN superiority (p < 0.001). These findings establish a new benchmark for classical approaches on GTZAN, demonstrating that traditional methods can achieve remarkable accuracy when combined with appropriate feature engineering and optimization strategies.