Session:

Knowledge Representation, Decision Support and Expert Systems
Tuesday, March 02, 2004, 12.30 – 12.50

Session Chair:
Vice Chair:

A. Dobnikar
M. Savoji

Paper Title:

MML-based Compressive Models for Musical Melody

Author(s):

A. Bickerstaffe, Monash University, Australia
Prof. D. L. Dowe, Monash University, Australia

Abstract:

Human inference of melodic structure is seemingly innate and intuitive, yet little is known about the cognitive processes that lead to such inference. Subsequently, computer modelling of melodic structure remains a difficult problem. Successful inference of the structure in musical data can provide insight into the process which created the data (e.g. the style of a composer) and result in data compression. Music is structurally rich data, with structure present at several levels, ranging from short term (e.g. note level) to long term (e.g. theme level). To systematically investigate melodic structure, we first begin with short term structure, that is, the structure of individual music notes. This paper provides eight new pitch models and one duration model for musical notes. These models are based upon the Minimum Message Length (MML) principle. Using MML, we discover which models best fit the test melodies and show that the best MML-based models compare favourably to existing compression techniques. We discuss limitations of the proposed methods and, finally, offer possible directions towards improving the MML-based models.