A Hierarchical Predictive Processing Approach to Modelling Prosody

Abstract

Prosodic patterns—and linguistic structures in general— are hierarchical in nature, providing for efficient means for encoding information in temporally constrained situations where communicative events occur. However, there are no theoretical frameworks that are capable of representing the full extent of linguistic behaviour in a cohesive way that could capture the paradigmatic and syntagmatic links between the organizational levels present in everyday speech. Here we propose a novel theoretical and modelling account of perception and production of prosodic patterns in speech communication, derived from the influential Predictive Processing theory of neural implementation of perception and action based on a hierarchical system of generative models producing progressively more detailed probabilistic predictions of future events. The framework provides a conceptualization of the hierarchical organization of speech prosody as well as a principled way of unifying speech perception and production by postulat-ing a single processing hierarchy shared by both modalities. We discuss the possible implications of the theory for prosodic analysis of speech communication, including conversational setting. In addition, we outline a viable computational implementation in the form of a machine learning architecture that can be used as a testbed for generating and evaluating predictions brought forth by the theory.