The Construction and Evaluation of Statistical Models of Melody and Harmony

Whorley, Raymond Peter. 2013. The Construction and Evaluation of Statistical Models of Melody and Harmony. Doctoral thesis, Goldsmiths, University of London [Thesis]

Text ( The Construction and Evaluation of Statistical Models of Melody and Harmony )
COM_thesis_Whorley_2013.pdf - Accepted Version

Download (11MB) | Preview

Abstract or Description

This research is concerned with the development of representational and modelling techniques employed in the creation of statistical models of melody and four-part harmony. Previous work has demonstrated the utility of multiple viewpoint systems, along with techniques such as Prediction by Partial Match, in the construction of cognitive models of melodic perception. Primitive viewpoints represent surface and underlying musical attributes, while linked viewpoints model combinations of such attributes. A viewpoint selection algorithm optimises multiple viewpoint systems by minimising the information theoretic measure cross-entropy. Many more linked viewpoints are used in this research than have previously been available, and the results show that many new viewpoints are incorporated into optimised systems.

A significant aspect of this work is the proposal and implementation of a set of novel extensions of the multiple viewpoint framework for four-part harmony. Statistical models are constructed with the aim that given a soprano part, alto, tenor and bass parts are added in a stylistically suitable way. Version 1 is as closely related to the modelling of melody as possible (chord replacing note), and is a baseline for gauging expected improvements as the framework is extended and generalised. Three versions of the framework have been implemented, and their performances compared and contrasted. The results indicate that the baseline version has been improved upon. Time complexity issues are discussed in detail, and selected viewpoints are examined from a music theoretic point of view for insights into why they perform well. Finally, melodies and harmonisations of given melodies are generated using the best performing models. The quality of the music suggests that, in spite of the improvements achieved so far, the models are still unable to fully capture the musical style of a corpus. Another six versions of the framework are described, which are expected to contribute further improvements.

Item Type:

Thesis (Doctoral)

Departments, Centres and Research Units:




Item ID:


Date Deposited:

27 Nov 2013 15:08

Last Modified:

08 Sep 2022 11:21


View statistics for this item...

Edit Record Edit Record (login required)