Extending a Model of Monophonic Hierarchical Music Analysis to Homophony

Computers are now powerful enough and data sets large enough to enable completely data-driven studies of Schenkerian analysis, the most well-established variety of hierarchical music analysis. In particular, we now have probabilistic models that can be trained via machine learning algorithms to analyze music in a hierarchical fashion as a music theorist would. Most of these models, however, only analyze the monophonic melodic content of the music, as opposed to taking all of the musical voices into account. In this paper, we explore the feasibility of extending a probabilistic model developed for analyzing monophonic music to function with homophonic music. We present details of the new model, an algorithm for determining the most probable analysis of the music, and a number of experiments evaluating the quality of the analyses predicted by the model. We also describe how varying the way the model interprets rests in the input music affects the resulting analyses produced.