Cemgil, T. and Kappen, B. (submitted) Bayesian Real-time Adaptation for Interactive Performance Systems
Abstract
An interactive music performance system (IMPS) is a computer program that listens to the actions of a performer and generates responses in real-time such as automatic accompaniment or improvisation (Rowe, 1993). One important goal is to design a robust IMPS that performs well for a broad set of performance conditions, e.g. different genres, styles, tempo e.t.c. One powerful machine learning strategy is statistical modeling, i.e. to devise a probabilistic model with adjustable parameters. Then, optimal parameters are estimated by maximization of the likelihood on a representative dataset. We have applied Bayesian online learning to a tempo tracking task. We compare the static model and the adaptive model by how well they predict the next beat in a given performance. On average, adaptation results in better predictions. For some performances the static filter is slightly better. However, for the majority of examples the prediction accuracy improves, and sometimes quite significantly. Interestingly, significant increases correspond to the a subject who uses consistently a lot of tempo variation. Static filter fails because personal optimal parameters of this particular subject are significantly different than other performers.
More on Time quantization