Self-Organizing Maps for Content-Based Music Clustering

Markus Frühwirth, Andreas Rauber
Department of Software Technology, Vienna University of Technology
Favoritenstr. 9 - 11 / 188, A-1040 Wien, Austria

Abstract:

With the increasing amount of music available electronically, methods for organizing these collections to allow intuitive browsing and orientation gain importance. Due to the large amounts of data involved, conventional approaches to organize music by genre or musical style are only of limited applicability, commonly relying on textual descriptions and manual classification. This makes it a particularly challenging application arena for neural networks capable of handling very high-dimensional input spaces and the noisy patterns associated with musical data.

In this paper we present a system based on the Self-Organizing Map which automatically organizes a collection of music files according to their musical genre and sound characteristics. Frequency spectra are used to extract feature vectors describing sound and melody characteristics. A two-stage clustering procedure first groups music segments according to their similarity, followed by a clustering of compositions according to the segment similarities. As a result, pieces of music with similar sound characteristics are found in neighboring regions of the resulting map, thus offering a very intuitive interface to unknown music collections.