Three Functions of Prediction Error for Bayesian Inference in Speech Perception

Spoken language is one of the most important sounds that humans hear, yet, also one of the most difficult sounds for non-human listeners or machines to identify. In this chapter we explore different neuro-computational implementations of Bayesian Inference for Speech Perception. We propose, in line with Predictive Coding (PC) principles, that Bayesian Inference is based on neural computations of the difference between heard and expected speech segments (Prediction Error). We will review three functions of these Prediction Error representations: (1) in combining prior knowledge and degraded speech for optimal word identification, (2) supporting rapid learning processes so that perception remains optimal despite perceptual degradation or variation, (3) ensuring that listeners detect instances of lexical novelty (previously unfamiliar words) so as to learn new words over the life span. Evidence from MEG and multivariate fMRI studies suggestion computations of Prediction Error in the Superior Temporal Gyrus (STG) during these three processes.