Modelling hierarchical musical structures with composite probabilistic networks

Abstract

The thesis is organised as follows:• Chapter 2 provides background information on existing research in the field of computational music harmonisation and generation, as well as some the¬ oretical background on musical structures. Finally, the chapter concludes with an outline of the scope and aims of this research.• Chapter 3 provides a short overview of the field of Machine Learning, ex¬ plaining concepts such as entropy measures and smoothing. The definitions of Markov chains and Hidden Markov models are introduced together with their methods of inference.• Chapter 4 begins with the definition of Hierarchical Hidden Markov models and techniques for linear time inference. It continues by introducing the new concept of Input-Output HHMMs, an extension to the hierarchical models that is derived from Input-Output HMMs.• Chapter 5 is a short chapter that shows the importance of the music rep¬ resentation and model structures for this research, and gives details of the representation.• Chapter 6 outlines the design of the software used for the HHMM modelling, and gives details of the software implementation and use.• Chapter 7 describes how dynamic networks of models were used for the generation of new pieces of music using a "random walk" approach. Several different types of networks are presented, exploring the different possibilities of layering the musical structures and organising the networks.• Chapter 8 tries to evaluate musical examples that were generated with sev¬ eral different types of networks. The evaluation process is both subjective and objective, using the results of a listening experiment as well as cross entropy measures and musical theoretical rules.• Chapter 9 offers a discussion of the methodology of the approach, the con¬ figuration and design of networks and models as well as the learning and generation of the new musical structures.• Chapter 10 concludes the thesis by summarising the research's contribu¬ tions, evaluating whether the project scope has been fulfilled and the major goals of the research have been met.