Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Non-semantic speech decoding in the hippocampus
Poster Session B, Friday, September 12, 4:30 - 6:00 pm, Field House
Yuchao Wang1, Aaron M. Earle-Richardson1, Derek Southwell1, Gerald Grant1, Muhammad Zafar1, Birgit Frauscher1, Gregory B. Cogan1; 1Duke University
The hippocampus has been recognized for its role in processing episodic and, more recently, semantic memory (Duff et al., 2020). However, evidence suggests that its function extends beyond memory to include speech and language processing, as demonstrated by hippocampal responses to sentence-level semantics (Piai et al., 2016), pronoun identity (Dijksterhuis, Self et al., 2024), and word boundaries (Ramos-Escobar et al., 2022). These responses to higher-level language processes may reflect overlaps with memory-related functions or indicate language and speech-specific processing. To further clarify the hippocampus’ role in speech processing, we investigated its involvement in speech perception and production. We collected data from intracranial electroencephalographic (iEEG) electrodes implanted in patients undergoing inpatient epilepsy monitoring for surgical resection at the Duke Comprehensive Epilepsy Center. In each trial, participants listened to an audio recording of a single word (or non-word; duration 0.5±0.1s) and spoke the word (or non-word) after a cue with 0.5-1.25s delay. Preprocessing included line noise filtering, muscle channel removal, outlier removal (>3SD), and common average referencing. The high gamma power (HG, 70-150Hz), which was shown to correlate with local neuronal firing, was then extracted using a multi-band band-pass filter. We first assessed whether there was word specific information in the hippocampus in Study 1 (N = 17, age = 33.6±12.5y, 7 Female). Of the 139 electrodes recorded in the hippocampus, 68 (49%) responded significantly during the listening or speaking epoch (cluster-corrected, time-resolved non-parametric shuffle test against baseline, FDR-corrected p < 0.05). We then trained a 4-way linear discriminant classifier with 5-fold cross validation to decode the target word from dimensionality-reduced HG activity (>80% variability retained after principal component decomposition) in a time-resolved manner (200ms window). The linear classifier achieved significantly above-chance decoding within 0.5s of onset in both listening and speaking (peak decoding = 36%, chance = 25%). This successful decoding suggests that the hippocampus processes speech related information. We then sought to validate and extend the findings to see if semantic information was necessary for hippocampal involvement in speech. In Study 2 (N=31, age = 30.9±10.9y, 16 Female), participants performed the same task with non-words. Using a similar decoding approach as the first experiment, we decoded the constituent phonemes within these non-words from neural signals in the hippocampus. Of 191 electrodes in the hippocampus, 66 (35%) responded significantly during listening or speaking. Our results showed significant decoding (peak decoding = 23%, chance = 11%), with a similar time-course of activation, providing further evidence for lower-level speech processing in the hippocampus. Together, these findings deepen our understanding of neural speech circuitry and establish the continuum of hippocampal involvement from sentence-level semantics to non-semantic phonemes to pure tones (Gillig et al., 2022).
Topic Areas: Speech Perception, Speech Motor Control