Bayesian Model Selection for Harmonic Labelling

We present a simple model based on Dirichlet distributions for pitch-class proportions within chords, motivated by the task of generating ‘lead sheets’ (sequences of chord labels) from symbolic musical data. Using this chord model, we demonstrate the use of Bayesian Model Selection to choose an appropriate span of musical time for labelling at all points in time throughout a song. We show how to infer parameters for our models from labelled ground-truth data, use these parameters to elicit details of the ground truth labelling procedure itself, and examine the performance of our system on a test corpus (giving 75% correct windowing decisions from optimal parameters). The performance characteristics of our system suggest that pitch class proportions alone do not capture all the information used in generating the ground-truth labels. We demonstrate that additional features can be seamlessly incorporated into our framework, and suggest particular features which would be likely to improve performance of our system for this task.