Predictive processing is a fundamental process in music cognition. While there are a number of predictive models of melodic structure, fewer approaches exist for harmony/chord prediction. This paper compares the predictive performance of n-gram, HMM, autoregressive HMMs as well as feature-based (or multiple-viewpoint) n-gram and Dynamic Bayesian Network Models of harmony, which used a basic set of duration and mode features. The evaluation was performed using a hand-selected corpus of Jazz standards. Multiple-viewpoint n-gram models yield strong results and outperform plain HMM models. However, feature-based DBNs outperform n-gram models and HMMs when incorporating the mode feature, but perform worse when duration is added to the models. Results suggest that the DBNs provide a promising route to modelling tonal harmony.

PDF file

In  Proceedings of the 9th International Symposium on Computer Music Modeling and Retrieval CMMR 2012

Publisher  Springer
Copyright remains with the authors

Details