Musical genre classification of audio signals using geometric methods

Musical genres are categorical labels characterizing pieces of music. Automatically classifying music into genres is gaining importance as a way to structure and organize the increasingly large numbers of music files available digitally on the web. In this work such a classification algorithm is developed and examined. The algorithm uses a vector of features based on the timbral texture of the music, and maps it into a new Euclidean space, by a non-linear method called “Diffusion Maps”, before the classification stage itself. This method allows dimensionality reduction while preserving and emphasizing the distinction between different genres. The proposed classifier classifies accurately 97% when classifying 2 musical genres, and 52% when classifying 10 musical genres. This is compared to an accuracy of 88% and 28% respectively, when classifying without the proposed mapping.