Measuring and modelling expression in piano performance

Systematic musicology employs concepts and methods from a wide range of parent disciplines, including mathematics, physics, computer science, biology, and the cognitive and social sciences. Musical expression involves structural, emotional and bodily aspects. We will focus on structural expression in piano music - the communication of musical structure by variations of timing, dynamics, articulation and pedalling. Sundberg and Friberg developed Director Musices, a system of context-dependent rules that introduces expressive deviation into automatic performances of input score files. We will combine this generative approach with the music-theoretical, accent-based approach of Parncutt (2003). An accent may be regarded as any event to which a listener`s attention may be attracted. Immanent accents may be deduced from the score and are associated with (temporal, serial) grouping (phrasing), metre (downbeats), melody (peaks, leaps) and harmony (or dissonance). Performed accents in piano music are generated by changes in timing, dynamics, articulation and pedalling, and vary in amplitude, form (amplitude as a function of time), and duration (the period of time during which the timing or dynamics are affected). We will ask how performed accents depend on immanent accents. For practical reasons, we will neglect other influences on musical expression. We plan to adjust Director Musices in two ways: conceptually, by introducing these accent-based concepts into the documentation, and mathematically, by adjusting algorithmic formulations, inspired by accent theory. The mathematical adjustments will be made by small steps that will be repeatedly checked by two empirical procedures: (i) comparison of generated performances (e.g. in MIDI format) with a databank of recorded performances, and (ii) perceptual experiments to evaluate the musical quality of generated performances. Richard Parncutt will supervise the project in Graz. Anders Friberg will assist with implementation of software changes. Gerhard Widmer, Luke Windsor, Caroline Palmer and Nicholas Cook have offered to contribute to our MIDI databank, against which generated performances will be compared. Widmer has also offered to compute tempo and loudness curves for specific recorded passages (Widmer, 2003). If the musical quality of performances generated by Director Musices can be improved by making clearer connections between music performance and music analysis, we will investigate the pedagogical implications in a follow-up project. Piano students will be trained to apply our approach to their performance, and their subsequent progress toward expressive goals will be documented.