Inducing Rules of Ensemble Music Performance : A Machine Learning Approach

Abstract

Previous research in expressive music performance has described how solo musicians intuitively shape each note in relation to local/global score contexts. However, expression in ensemble performances, where each individual voice is played simultaneously with other voices, has been little explored. We present an exploratory study in which the performance of a string quartet is recorded and analysed by a computer. We use contact microphones to acquire four audio signals from which a set of audio descriptors is extracted individually for each musician. Moreover, we use motion capture to extract bowing descriptors (bow velocity/force) from each of the four performers. The gathered multimodal data is used to align the performance to the score. Then, from the aligned data streams, we obtain a note-by-note description of the performance by extracting note performance parameters. We apply machine-learning algorithms to induce human-readable rules emerging from the data. The dataset consists of three performances of Beethoven’s quartet n° 4 in C minor by a group of professional musicians: a “normal”, a “mechanical” and an “over-emphasized” execution. We run our analysis on the three conditions separately as well as jointly, deriving rules specific to each condition and rules of general domain. Apart from encoding knowledge of expressive performance, the results shed light on how musicians' roles in ensemble performance.