Experienced listeners Modeling the influence of long-term musical exposure on rhythm perception

This thesis investigates rhythm perception using computational modeling techniques. It develops a set of tools and techniques for the definition and evaluation of probabilistic generative models of music perception. It argues that previously proposed computational models of rhythm perception insufficiently account for the shaping of rhythm perception by prior experiences, practice, and training of listeners. Motivated by predictive processing theories of perception, the thesis proposes a probabilistic generative model of meter perception which, compared with previous models, to a greater extent learns from regularities in datasets of rhythms (representing a musical environment). The outcome of this learning process simulates the effects of the long-term exposure to rhythms in the musical environment of listeners. The model is introduced via a novel framework, introduced in this thesis, in which probabilistic generative models of music cognition can be formulated as dynamic Bayesian networks with deterministic constraints. The thesis furthermore presents cross-cultural simulations in which the novel model introduced in this thesis is compared to a model that represents classical theories of rhythm perception on two datasets that respectively represent rhythms from Western folksongs and Turkish makam music. The results show that both datasets contain regularities that allow the novel model but not the model representing classical theories to infer meter from individual rhythms. The results furthermore show that some of these regularities are specific to Western or Turkish rhythms, although regularities by which meter can be inferred also appear to be shared to a significant extent between Turkish and Western rhythms.