The present work describes a system for the automatic recognition of music genres, based exclusively on the audio content of the signal. A comprehensive overview of music genre classification is presented, including disquisitions on the required background from fields such as human sound perception, pattern classification, and signal processing. A critical evaluation of the tacit assumptions behind many existing approaches is made, resulting in a number of conclusions about the design of a music genre recognition system. A prototypical system ( MUGRAT) that is based on these principles is presented.

MUGRAT is based on extracting various features from the input sound that are likely to also be important in human music genre recognition. These features can roughly be distinguished into two categories: (a) features related to the musical texture, and (b) features related to the rhythm/beatedness of the sound. A k-nearest-neighbour classifier is trained with a set of sample data consisting of randomly chosen musical excerpts of 3 seconds length. The system reaches a classification accuracy of 88 % for the three test genres Metal, Dance, and Classical.