Unsupervised Classification of Music Signals: Strategies Using Timbre and Rhythm

This thesis describes the ideal properties of an adaptable music classification system based on unsupervised machine learning, and argues that such a system should be based on the fundamental musical properties of timbre, rhythm, melody and harmony. The first two properties and the signal features associated with them are then explored in more depth. In the area of timbre, the relationship between musical style and commonly-extracted signal features within a broad range of piano music is explored, in an effort to identify features which are consistent among all piano music but different for other instruments. The effect of lossy compression on these same timbre features is also investigated. In the area of rhythm, a new tempo tracking tool is provided which produces a series of histograms containing beat and sub-beat information throughout the course of a musical recording. These histograms are then shown to be useful in the analysis of synthesized rhythms and real music. Additionally, a novel method based on the Expectation-Maximization algorithm is used to extract features for classification from the histograms.