Towards Seamless Network Music Performance: Predicting an Ensemble's Expressive Decisions for Distributed Performance

Internet performance faces the challenge of network latency. One proposed solution is music prediction, wherein musical events are predicted in advance and transmitted to distributed musicians ahead of the network delay. We present a context-aware music prediction system focusing on expressive timing: a Bayesian network that incorporates stylistic model selection and linear conditional gaussian distributions on variables representing proportional tempo change. The system can be trained using rehearsals of distributed or co-located ensembles. We evaluate the model by comparing its prediction accuracy to two others: one employing only linear conditional dependencies between expressive timing nodes but no stylistic clustering, and one using only independent distributions for timing changes. The three models are tested on performances of a custom-composed piece that is played ten times, each in one of two styles. The results are promising, with the proposed system outperforming the other two. In predictable parts of the performance, the system with conditional dependencies and stylistic clustering achieves errors of 15ms; in more difficult sections, the errors rise to 100ms; and, in unpredictable sections, the error is too great for seamless timing emulation. Finally, we discuss avenues for further research and propose the use of predictive timing cues using our system.