Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Lexical surprisal modulates primary auditory cortex during naturalistic listening
Poster Session F, Friday, October 2, 2:45 - 4:45 pm, Wangari Maathai
Pan Zhang1, Alejandro Tabas1,2,3,4; 1Basque Center on Cognition, Brain and Language, 2Ikerbasque, Basque Foundation for Science, 3Max Planck Institute for Human Cognitive and Brain Sciences, 4Technische Universität Dresden
Humans are remarkably proficient at understanding speech in noisy contexts. Recent evidence showed that the brain predicts what the speaker may say next to aid such challenging processing. But how far down the processing hierarchy are these predictions used? Prior work has consistently localised predictive effects to superior temporal gyrus, inferior frontal gyrus, and anterior temporal lobe, concluding that prediction is confined to a fronto-temporal “language network”. This view portrays earlier areas such as the primary auditory cortex (PAC) as upstream acoustic relaying stations. However, for rapid and noisy signals such as speech, predictions can be extremely useful even at early stages where features need to be detected under high noise. Here we explore the reach of high order lexical predictions in early auditory regions. We reasoned that, if predictions aid the speech processing, word predictability should be inversely related to processing effort and hence to the amplitude of the neural responses elicited. We used the Nastase “Narratives” fMRI dataset to test whether predictability affected BOLD responses in six bilateral auditory ROIs, i.e., anterior Heschl’s gyrus (aHG), planum temporale (PT), and lateral STG (La). The dataset includes whole-brain fMRI measurements of 345 participants listening to natural speech (TR = 1.5s, 2.0/2.5/3.0mm isotropic voxels). We built a model assuming that responses were modulated by predictability with two parametric modulators: one encoding expected responses to the acoustic properties (acoustic envelope) and one encoding word-level surprisal (the negative log-probability of the word given contexts, estimated by GPT-2). We constructed twenty independent null models, with the same acoustic regressors as the main model and random permutations of the surprisal. We then computed group-level Bayes factors (K) comparing the acoustics+surprisal model against the acoustics+shuffled-surprisal null models. In addition, to study the relative strength of the modulation exerted by surprisal in each ROI, we computed an index encoding the relative contribution of the acoustics and surprisal modulators of the main model. Distributions of these indices were compared between ROIs using paired Wilcoxon signed-rank tests (FDR-corrected). According to Jeffery’s scale, evidence for modulation of the responses by lexical practicability was substantial in left aHG (logK>1.15) and moderate in right aHG (logK>1.06); strong in PT (left: logK>1.33); and decisive in La (left: logK>4.45). We found gradient and lateralised effects of the relative contribution of surprisal consistent with previous literature. Predictive modulation effects increased from aHG to PT to La (d > 0.47, p < 10^-14) and were stronger in the left hemisphere (d > 0.20, p < 10^-5). There are three main findings. First, word-level modulation lexical predictability is apparent as early as in aHG. Second, the modulation effect strengthens along the auditory hierarchy, extending the previously documented cortical gradient to PAC. Third, the predictive modulation is consistently left-lateralised. Together, these results demonstrate that lexical predictions are already used to aid processing in PAC, framing the hierarchy from PAC to lateral STG as a graded, left-lateralised predictive pathway. This intricate communication of predictions turns the challenging computation of everyday speech into an effortless experience.
Topic Areas: Speech Perception,