Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Single neuron encoding of acoustic and linguistic information in the human hippocampus during story listening
Poster Session D, Thursday, October 1, 4:30 - 6:30 pm, Wangari Maathai
Corentin Puffay1,3, Melissa Franch2, Benjamin Hayden2, Laura Gwilliams3; 1KU Leuven, 2Baylor College of Medicine, 3Stanford University
Introduction. Over the past decade, studies have elucidated the role of the hippocampus in processing concepts: Finding evidence for hippocampal neurons that encode a concept such as “Luke Skywalker” regardless of sensory modality (Quiroga et al., 2009). Building on these findings, recent work suggests that the hippocampus also contributes to conceptual and semantic processing during language comprehension, by encoding word meaning during story listening (Franch et al., 2026). At the neocortical level, language processing has been widely investigated, providing evidence for the emergence of a hierarchy of acoustic and linguistic representations, distributed across the language network (Friederici et al., 2011; Gwilliams et al., 2025). At the allocortical level, on the other hand, no such characterization has yet been established, raising questions about the underlying mechanisms of language processing beyond the neocortex. To bridge this gap, here we address the question: Is language encoding in the hippocampus restricted to word meaning, or are multiple levels of acoustic and linguistic representations encoded during continuous, naturalistic speech processing? Methods. We analyzed neuronal spike activity from 10 epilepsy patients implanted with Behnke-Fried electrodes who listened to six podcast narratives. We fit linear encoding models relating neural firing to multiple features spanning different levels of the speech hierarchy: low-level acoustics (speech envelope), lexical structure (word and sentence onset timings), and higher-level linguistic (word surprisal and contextual embeddings derived from GPT2). To characterize both feature selectivity and temporal structure of neural responses, we performed variance partitioning analyses and estimated temporal response functions (TRFs) for individual neurons. This approach allowed us to quantify how different speech features are mapped onto hippocampal activity over time, capturing both their unique encoding strengths and latency profiles. Results. We found that hippocampal neurons tune to different speech features, with neurons preferentially encoding either acoustic or linguistic information. Importantly, these representations also differed in their temporal dynamics. Acoustic features were associated with earlier, more transient neural responses, whereas linguistic and semantic features were associated with delayed, more sustained neural responses. Conclusion. Together, these findings suggest that neurons in the human hippocampus not only encode concepts and word meanings, but rather multiple levels of speech information during naturalistic listening. Acoustic and linguistic representations are both encoded by hippocampal neural populations, with distinct temporal profiles. This extends hierarchical models of speech processing beyond the neocortex and supports a role for the hippocampus in representing structured linguistic information at single-neuron resolution during continuous speech.
Topic Areas: Speech Perception, Computational Approaches