SYSTEMATIC EVALUATION AND IMPROVEMENT OF STATISTICAL MODELS OF HARMONY

We are investigating the utility of Markov models in relation to the learning of the task of four-part harmonisation, which is a creative musical activity. A program is described which uses statistical machine learning techniques to learn this task from a suitable corpus of homophonic music. The task is decomposed into a series of more manageable sub-tasks; these are each modelled by Markov models, which can use contexts drawn from symbols describing past, current and future chords. The results of a number of initial studies, for example comparing different types of model and the effect of corpus size, are given. There is also some discussion about harmonisations that have been generated by the program by random sampling of the probability distributions in the models. Following this, a procedure for the systematic evaluation and “optimisation” of the sub-task models, involving the application of an information-theoretic measure, is presented, along with some more results. An appraisal of the procedure’s shortcomings is made, and ideas for its improvement are put forward. Finally, an indication of the future direction of the work (which is currently in its early stages) is given.