MML classification of music genres

Abstract

Inference of musical genre, whilst seemingly innate to the human mind, remains a challenging task for the machine learning community. Online music retrieval and automatic music generation are just two of many interesting applications that could benefit from such research. This paper applies four different classification methods to the task of distinguishing between rock and classical music styles. Each method uses the Minimum Message Length (MML) principle of statistical inference. The first, an unsupervised learning tool called Snob, performed very poorly. Three supervised classification methods, namely decision trees, decision graphs and neural networks, performed significantly better. The..