Markov Random Fields and Maximum Entropy Modeling for Music Information Retrieval

Music information retrieval is characterized by a number of various user information needs. Systems are being developed that allow searchers to find melodies, rhythms, genres, and singers or artists, to name but a few. At the heart of all these systems is the need to find models or measures that answer the question "how similar are two given pieces of music". However, similarity has a variety of meanings depending on the nature of the system being developed. More importantly, the features extracted from a music source are often either single-dimensional (i.e.: only pitch, or only rhythm, or only timbre) or else assumed to be orthogonal. In this paper we present a framework for developing systems which combine a wide variety of non-independent features without having to make the independence assumption. As evidence of effectiveness, we evaluate the system on the polyphonic theme similarity task over symbolic data. Nevertheless, we emphasize that the framework is general, and can handle a range of music information retrieval tasks