Cortical tracking of prediction error during perception of connected speech

Abstract

Prediction facilitates speech comprehension, but how predictions are combined with sensory input during perception remains unclear. Prior studies suggest that prediction error computations -- which represent the difference between heard and expected speech sounds -- play a central role in speech perception. However, these studies have relied on non-ecological listening conditions, involving isolated words predicted by artificial cues. In two experiments, we presented 61 listeners with semantically coherent sentences that ended in words of varying predictability. We also manipulated the signal quality of the final word by applying time-limited degradation. Using linear encoding models of EEG responses, we found that neural representations of speech features were jointly influenced by top-down predictions and bottom-up signal quality. Specifically, for unpredicted final words, increasing signal quality resulted in enhanced neural tracking of speech acoustic modulations in low frequency EEG (0-2 Hz delta band). In contrast, for strongly predicted final words, greater signal quality led to reduced tracking of speech modulations. Computational simulations revealed that this interaction is consistent with prediction error computations, but not with an alternative signal sharpening model. These findings extend the evidence for prediction error computations to more naturalistic listening situations.