MACHINE LEARNING TECHNIQUES APPLIED TO MUSICAL GENRE RECOGNITION

Abstract

With the wealth of music available at the fingertips of users around the world, there isan ever-increasing need for automatic classification of music for cataloguing of music fororganization and quicker retrieval which is often done manually by experts in the field. Tofurther complicate the issue, there is no standard definition on what determines a song’sgenre, which can be a culmination of various themes and moods that the song generates inlisteners. This work designs and evaluates several models using Deep Neural Networks andGradient Boosting Machines, using various transformations of the raw audio for predictingthe genre of a particular piece of music. In particular, this research involves adapting thenatural taxonomy of musical genres to generate a machine learning model in an attemptto capture some of the natural hierarchy in music. The results show that gradient boostingmachines outperform all other models in terms of loss and accuracy.