Analysis and prediction of expressive dynamics using Bayesian linear models

We present the first version of a probabilistic linear basis model for expressive dynamics in music. The model extends prior work by Grachten and Widmer [7]. Contrary to the original model, this model allows for both specifying musical knowledge, and for modeling multiple distinct performances of the same piece. We show that in its current, minimalist form, the new model performs on a par with the original model in terms of predictive accuracy. Furthermore, a novel set of basis-functions is evaluated, to model dynamics annotations such as (de)crescendo in a context-aware manner.