Rapid computations of spectrotemporal prediction error support perception of degraded speech

Abstract

Human speech perception can be described as Bayesian perceptual inference but how are these Bayesian computations instantiated neurally? We use magnetoencephalographic recordings of brain responses to degraded spoken words as a function of signal quality and prior knowledge to demonstrate that spectrotemporal modulations in speech are more clearly represented in neural responses than alternative speech representations (e.g. spectrogram or articulatory features). We found an interaction between speech signal quality and expectations from prior written text on the quality of neural representations; increased signal quality enhanced neural representations of speech that mismatched with prior expectations, but led to greater suppression of speech that matched prior expectations. This interaction is a unique neural signature of prediction error computations and already apparent in neural responses within 250 ms of speech input. Our findings contribute towards the detailed specification of a computational model of speech perception based on predictive coding frameworks.