The neural representation of missing speech and the influence of prior knowledge on cortical fidelity and latency

Abstract

In naturally noisy listening conditions, for example at a cocktail party, noise disruptions may completely mask significant parts of a sentence, and yet listeners may still perceive the missing speech as being present. Here we demonstrate that dynamic speech-related auditory cortical activity, as measured by magnetoencephalography (MEG), which can ordinarily be used to directly reconstruct to the physical speech stimulus, can also be used to \"reconstruct\" acoustically missing speech. The extent to which this occurs depends on the extent that listeners are familiar with the missing speech, which is consistent with this neural activity being a dynamic representation of perceived speech even if acoustically absence. Our findings are two-fold: first, we find that when the speech is entirely acoustically absent, the acoustically absent speech can still be reconstructed with performance up to 25% of that of acoustically present speech without noise; and second, that this same expertise facilitates faster processing of natural speech by approximately 5 ms. Both effects disappear when listeners have no or very little prior experience with a given sentence. Our results suggest adaptive mechanisms of consolidation of detailed representations about speech, and the enabling of strong expectations this entails, as identifiable factors assisting automatic speech restoration over ecologically relevant timescales.