Duet interaction: learning musicianship for automatic accompaniment

Computer music systems can interact with humans at different levels, including scores, phrases, notes, beats, and gestures. However, most current systems lack basic musical skills. As a consequence, the results of human-computer interaction are often far less musical than the interaction between human musicians. In this paper, we explore the possibility of learning some basic music performance skills from rehearsal data. In particular, we consider the piano duet scenario where two musicians expressively interact with each other. Our work extends previous automatic accompaniment systems. We have built an artificial pianist that can automatically improve its ability to sense and interact with a human pianist, learning from rehearsal experience. We describe different machine learning algorithms to learn aspects of expressive timing and dynamics for duet interaction, explore the properties of the learned models, such as dominant features, limits of validity, and minimal training size, and claim that a more human-like interaction is achieved.