Prior Context Scaffolds Sentential Semantic Integration during Noisy Speech Comprehension

Abstract

Understanding speech in noise is a central challenge of everyday communication, yet listeners often succeed by using prior context. How the brain uses such context remains debated: it may refine predictions about upcoming words, or it may provide a higher-level framework that helps degraded speech cohere into meaning. Here we combined simultaneous EEG-fNIRS recording with hierarchical multivariate encoding models to track how prior context shapes speech processing from acoustics to words and sentential meaning. Participants listened to natural spoken narratives under clear speech, noisy speech, and context-supported noisy speech conditions. Context brought comprehension of noisy speech close to clear-speech levels. EEG revealed that contextual support reduced neural encoding of lexical surprisal and entropy, indicating weaker tracking of local word-level prediction demands. In contrast, when context was available, fNIRS showed enhanced encoding of sentence-level semantic integration across frontal regions and the right angular gyrus, and stronger angular gyrus encoding predicted better comprehension. By combining EEG and fNIRS to capture complementary electrophysiological and hemodynamic signals, this multimodal approach reveals a hierarchical shift in degraded speech comprehension: prior context does not simply improve word-by-word prediction, but scaffolds the integration of noisy input into coherent discourse.