Dissecting expression into structural components

Peter Desain1, Luke Windsor² & Amandine Penel³

1Radboud University Nijmegen, the Netherlands ²University of Leeds, UK ³Cold Spring Harbor Laboratory, USA

A method for separating, profiling and quantifying the contributions of different structural components to expressive musical performance is described. The (so-called DISSECT) method is demonstrated through its application to a set of expert piano performances of a short piece from the classical period. It uses a generalized linear regression model based on a structural annotation of the score. In contrast to existing models this method can deal with many (possibly incompatible) types of structural descriptions, like metrical, phrase, and local. The model succeeds in explaining quite a large percentage of the variance, and aids in the understanding how the different structural components in a piece of music combine in the generation of an expressive performance. Though in the present study some simplifications have been made, like ignoring the asynchrony between voices and grace note timing, these aspects can be analyzed with the same means. Also the effect of global tempo on the various expressive components can be studied in detail. The method is compared with other approaches to the analysis and modeling of musical performance. Potential applications are identified, one of which is the possibility of expression mixing: the reconstitution of a fully expressive performance with parameters controlling for each structural element to what extend it is expressed in timing and possibly dynamics.