Automatic subgenre classification of heavy metal music

Abstract

Automatic genre classification of music has been of interest for researchers over a decade. Many success-ful methods and machine learning algorithms have been developed achieving reasonably good results. This thesis explores automatic sub-genre classification problem of one of the most popular meta-genres, heavy metal. To the best of my knowledge this is the first attempt to study the issue. Besides attempting automatic classification, the thesis investigates sub-genre taxonomy of heavy metal music, highlighting the historical origins and the most prominent musical features of its sub-genres. For classification, an algorithm proposed in (Barbedo & Lopes, 2007) was modified and implemented in MATLAB. The obtained results were compared to other commonly used classifiers such as AdaBoost and K-nearest neighbours. For each classifier two sets of features were employed selected using two strategies: Correlation based feature selection and Wrapper selection. A dataset consisting of 210 tracks representing seven genres was used for testing the classification algorithms. Implemented algorithm classified 37.1% of test samples correctly, which is significantly better performance than random classification (14.3%). However, it was not the best achieved result among the classifiers tested. The best result with correct classification rate of 45.7% was achieved by AdaBoost algorithm.