Gaussian Process Regression for Rendering Music Performance
Keiko Teramura, Hideharu Okuma, Yuusaku Taniguchi, Shimpei Makimoto, Shin‐ichi Maeda · 2008
So far, many of the computational models for rendering music performance have been proposed, but they often consist of many heuristic rules and tend to be complex. It makes difficult to generate and select the useful rules, or perform the optimization of parameters in the rules. In this study, we present a new approach that automatically learns a computational model for rendering music performance with score information as an input and the corresponding real performance data as an output. We use a Gaussian Process (GP) incorporated with a Bayesian Committee Machine to reduce naive GP's heavy computation cost, to learn those input-output relationships. We compared three normalized errors: dynamics, attack time and release time between the real and predicted performance by the trained GP to evaluate our proposed scheme. We evaluated the learning ability and the generalization ability. The results show that the trained GP has an acceptable learning ability for 'known' pieces, but show insufficient generalization ability for 'unknown' pieces, suggesting that the GP can learn the expressive music performance without setting many parameters manually, but the size of the current training dataset is not sufficiently large so as to generalize the training pieces to 'unknown' test pieces.