Music Genre Classification Application Based on Audio Features with Ensemble Learning Algorithm

Abstract

With the exponential growth of digital music, manual genre-labeling has become ineffective. Consequently, automatic music genre classification is crucial for data management and recommendation systems. This research aims to develop an accurate music classification application by comparing individual machine learning models against advanced Ensemble Learning techniques. The methodology involved extracting 26 audio features from the GTZAN dataset, followed by training and hyperparameter tuning ten models, including Random Forest, SVM, XGBoost, and LightGBM. The findings demonstrate that ensemble methods significantly outperform individual models. The highest performance was achieved by a Voting Classifier, which combines the predictive strengths of SVM, XGBoost, and Logistic Regression, reaching a final test accuracy of 72%. This superior ensemble model was then successfully implemented into an interactive web application using Streamlit, proving that this approach is not only highly accurate but also functional for real-time, practical applications.