ABSTRACT

This chapter reviews some qualitative aspects of the trained recurrent nonlinear basis model (RNBMs), revealing some of the temporal dependencies the model has learned. It focuses on a few data-driven models that are intended to infer regularities in musical expression from datasets of recorded performances. G. Grindlay and D. Helmbold describe a hidden Markov model (HMM) of expressive timing. The model involves a number of hidden states, where each state is associated with a joint probability distribution over the score information and the expressive timing. The chapter describes the "linear versus nonlinear basis models" and "static versus temporal basis models". The results of these models seem to confirm the respective benefits of nonlinearity and temporal dependencies for modeling expressive timing in classical piano performances. The chapter analyzes the effects of past and future score information on the performance of a note, using a method called differential sensitivity analysis.