Visual speech enhances phoneme separability in human superior temporal gyrus

Abstract

Visual speech, such as lipreading, facilitates spoken word recognition, but the neural mechanisms underlying audiovisual speech perception remain poorly understood. Visual cues may disambiguate fine-grained articulatory features during early perceptual stages or instead integrate with speech at more categorical, phoneme-level stages. To test how speech representations are modulated by visual input, we analyzed intracranial electroencephalography (iEEG) signals recorded from 12 epilepsy patients performing an audiovisual speech perception task. Participants perceived 16 monosyllabic words presented in auditory-only, visual-only, or congruent audiovisual formats. Words were constructed from four onset consonants (/b/, /g/, /m/, /n/) and four rimes (vowel nucleus and any coda consonants). We examined event-related potentials (ERP) in superior temporal gyrus (STG) and trained support vector machine (SVM) classifiers to decode word identity from neural activity at individual electrodes. Discrete and continuous confusion matrices captured complementary changes in classification accuracy and normalized inverse classification loss, a continuous proxy for classifier confidence. Decoding performance was hierarchically evaluated at the word, phoneme, and phonetic feature levels to determine the representations affected by visual speech. Congruent audiovisual speech increased classifier confidence for phoneme-level representations and improved decoding accuracy at both the word and phoneme level, without corresponding effects on phonetic features. Time-resolved analyses further revealed earlier successful decoding for audiovisual than auditory-only speech, with audiovisual enhancement primarily observed for onset consonants rather than rimes. Together, these findings suggest that visual speech sharpens primarily categorical phoneme representations in STG, with accelerated speech processing and improved word recognition emerging as downstream consequences of phoneme-level enhancement. Significance Statement Seeing a speaker’s face improves speech perception, especially in noise, but the neural representations altered by visual speech remain unclear. Using intracranial recordings from human superior temporal gyrus, we show that congruent visual speech does not simply amplify all speech-related information. Instead, visual input selectively enhances the separability of confusable phoneme-level representations and improves word identity decoding, while leaving the robust phonetic-feature organization largely intact. These findings clarify the distinction between what visual speech modulates and the representational structure of speech in auditory cortex, suggesting that audiovisual facilitation primarily acts on categorical speech representations that more directly support word recognition.