Learning to annotate music files using content based retrieval systems and wavelet packet approximations of the input signals

Abstract

This thesis presents an inter-disciplinary approach to the problem of music parametrisation for information representation and retrieval. Signal analysis and machine learning techniques are combined in the context of music characterisation. The vast amount of music available electronically presents considerable challenges for information retrieval. This research focuses on k-nearest neighbour classifiers for content base retrieval of music files. Classifiers that support query by example are presented and evaluated. The thesis presents a process for determ ining the genre of a music file using a new set of descriptors. A recent technique based on the wavelet transform is the core technique used for characterisation: a discrete wavelet packet transform is applied to obtain the signal representation at different decomposition levels. Time and frequency features are extracted from these levels taking into account the specific requirements of music parametrisation; its time and frequency constraints.