Tracking melodic expectations during naturalistic music listening using M|EEG

Abstract

Expectations are widely thought to shape our experience of music. However, neural evidence mainly stems from studies employing constrained paradigms and stimuli, which challenges the predictive nature of naturalistic music listening. Furthermore, it is debated which sources and temporal context melodic expectations rely on. In the current study, we presented human listeners with naturalistic monophonic compositions from Western classical music, while recording neural activity using MEG. We quantified note-level melodic surprise and uncertainty via computational models of music, resembling different sources of melodic expectations: Gestalt principles, short-term regularities or long-term statistical learning. For the first time, we applied a state-of-the-art neural network, the Music Transformer, to study music cognition. We demonstrate its sensitivity to long-range musical structure compared to previous models, which allowed us to probe the influence of musical context length on neural music processing. A time- resolved regression analysis revealed that melodic surprise increased neural responses over fronto-temporal areas particularly around 200 ms and 300–500 ms after note onset, which was dissociated from sensory-acoustic and repetition suppression effects. According to a model comparison on cross-validated predictive performance, neural surprise was best captured by long-term statistical learning and short-range musical contexts of less than ten notes. Uncertainty, on the other hand, was not found to modulate neural activity evoked by notes. All results were confirmed on a recently published EEG dataset (Di Liberto et al., 2020). Our findings provide evidence for the role of melodic expectations during naturalistic music listening, that emerge from musical enculturation.

Keywords: Music listening, naturalistic, computational modelling, prediction, statistical learning