A Genetic Rule-Based Model of Expressive Performance for Jazz Saxophone

摘要

This article describes an evolutionary-computation approach for learning an expressive-performance model from recordings of jazz standards by a skilled saxophone player. Our objective has been to find a computational model that predicts how a particular note in a particular context should be played (e.g., longer or shorter than its nominal duration). To induce the expressive-performance model, we extracted a set of acoustic features from the recordings resulting in a symbolic representation of the performed pieces, and we then applied a sequential-covering genetic algorithm to the symbolic data and information about the context in which the data appear. Despite the relatively small amount of training data, the induced model seems to accurately capture the musician's expressive-performance transformations. In addition, some of the classification rules induced by the algorithm proved to be of musical interest. Currently, we are in the process of increasing the amount of training data as well as experimenting with different information encoded in the data. Increasing the size of the training data set, extending the information in it, and combining it with background musical knowledge will certainly generate more models. We are also extending