Real-time Piano Accompaniment Model Trained on and Evaluated According to Human Ensemble Characteristics

With a view towards the goal of modelling the performance of classical music by human musicians, we have set our focus on the question of collaborative music-making in a MIDI environment. In previous work, we have presented a model that plays a part of a score in real time together with a live musician playing another part. We trained it to resemble human musicians faced with the same task, by tuning its systems built around a set of Kuramoto oscillators. Here we chose 3 musical works and conducted experiments collaborating with a variety of pianists and record the resulting performances as well as the testers’ subjective impressions. We reconciled each performance with the corresponding music score, thereby defining a dataset which we call an "interpretation". In addition to subjective evaluation, we introduced objective criteria in the form of discriminants that classify interpretations as being the result of human-human interaction or of human-machine interaction. We considered the following qualities: desynchronization, jerkiness, and velocity curves. Our trained model performed similarly to humans with respect to the first two discriminants, but significantly differently with respect to the last. In light of this, it is notable that our experiment subjects often failed to correctly distinguish the two classes.