Orchestral Accompaniment for a Reproducing Piano

A system that generates flexible orchestral accompaniment of a computer-enabled piano is presented and demonstrated. We introduce a probabilistic model for the piano data that can be used for on-line and off-line estimation of the piano performance. The model is automatically trainable to the specific performer and piece under consideration. The on-line position estimates form the observable data for a trainable prediction engine that anticipates the future evolution of the performance. These ongoing predictions drive a phase-vocoded audio performance of the orchestra. We present results on a highly challenging gem from the Romantic piano concerto repertoire.