Stochastic analysis of music

In this talk I will explore the usefulness of stochastic parsing techniques for predicting musical grouping structure. One of the main challenges in modeling grouping structure is the problem of ambiguity: many different structures are compatible with a musical piece while a listener typically perceives only one structure. In this talk I consider a number of parsing techniques that use a probabilistic approach to solve ambiguity. Rather than by a predefined set of rules or principles, probabilistic approaches learn to predict the perceived grouping structure by generalizing over a set of previously annotated pieces. I will discuss some experiments with the Essen Folksong Collection and I will give an in-depth comparison between probabilistic and rule-based models of grouping structure. It turns out that there are grouping phenomena (related to so-called jump-phrases) that challenge the commonly accepted Gestalt principles of proximity and similarity, and that cannot be explained by other musical factors such as meter or harmony. I conjecture that probabilistic, memory-based approaches are more apt to modeling grouping structure since they mimic the musical experience of a listener from a certain culture. Time-permitting, I will also go into a number of recent psychological experiments carried out at the University of Amsterdam that seem to support my conjecture.