Mathematical Models of Musical Performance

Trained musicians intuitively add expressive variations to the music they perform. These changes in tempo and loudness are the foundations of a compelling performance. However, there is little quantitative information about the kinds of strategies employed by different musicians and in different musical contexts. This situation provides an opportunity for machine learning techniques that can model the statistically salient aspects of musical performance and automatically produce compelling expressive variations. We are investigating new methodologies for capturing these expressive variations in performance-specific models. The resulting system can be used to synthesize performances of new pieces of music in the style of particular performers. Furthermore, the generative models learned by the system are useful for search and classification applications and also may provide insight into a number of musicological questions related to expressive performance.