Representation of musical performance “grammar” using probabilistic graphical models

Abstract

As a versatile data modeling tool, probabilistic graphical model can be applied to model the complex dependency structures encoded in the contextual “grammar” of music performance. The musical performance grammar here refers to the relational structures of the sonic features extracted from music performances. In the existing literature, the data structure of musical expressive grammar is usually modeled as rule list, following the grammatical format of natural language processing applications. In this work, we apply the representation format of probabilistic graphical model to musical performance features to extend the conventional rule-list format. We choose probabilistic graphical model as an “upgraded” representation for two reasons. First, probabilistic graphical model provides enhanced representation capability of relational structures. This feature enables us to model the complex dependency structure that the conventionally rule list cannot handle. Second, the graphical format of probabilistic graphical model provides an intuitive human-data interface and allows in-depth data visualization, analysis, and interaction. We include the representation and analysis examples of musical performance grammar obtained from both manual analysis and automatic induction. We also implemented interpretation tools that interface the rule-list format and the probabilistic graphical model format to enable detailed comparison with existing results of musical performance analysis.