Poster Presentation

Search Abstracts | Symposia | Slide Sessions | Poster Sessions

Representations spanning phonemes to syntax during speech production at single-neuron resolution

Poster Session E, Friday, October 2, 11:00 am - 1:00 pm, Wangari Maathai

Alessandro Marin Vargas¹, Seonghyun Yoon¹, Alisa D. Levin¹, Sasidhar Madugula¹, Erin Kunz¹, Ryan Wang¹, Chaofei Fan¹, Qiqi Xian¹, Akansha Singh¹, Christina Vo¹, Donald T. Avansino¹, Payton Bechefsky^2,3, Samuel R. Nason-Tomaszewski^2,3, Yuhan Zhang¹, Erxiao Wang¹, Leigh R. Hochberg^4,5,6, Nicholas AuYong^2,3, Chethan Pandarinath^2,3, Cory Shain¹, Jaimie Henderson¹, Francis R. Willett¹; ¹Stanford University, ²Emory University, ³Georgia Institute of Technology, ⁴Brown University, ⁵VA Providence Healthcare System, ⁶Massachusetts General Hospital, Harvard Medical School

Prior fMRI and ECoG/sEEG studies have linked distinct cortical regions to different levels of the linguistic hierarchy, but how these features are represented at single-neuron resolution remains poorly understood. Here, using intracortical recordings obtained from people with paralysis in the BrainGate2 pilot clinical trial, we asked how motoric and linguistic features are represented during speech planning and production at single-neuron scale. Consistent with prior work, we first identified a clear anatomical dissociation: arrays in ventral precentral gyrus (area 6v) were preferentially explained by phonemic and motor-related features, whereas arrays in language-related frontal sites in middle frontal gyrus and inferior frontal gyrus showed stronger tuning to abstract linguistic features. We then used encoding models to determine which levels of linguistic structure best explained activity in language-related regions. Using a large sentence-repetition dataset (131k words), we fit encoding models based on lexical, syntactic, semantic and LLM-derived contextual representations and quantified how well each feature space explained neural activity. We found relatively weak tuning to word-level lexical and semantic features, but substantially stronger tuning to syntactic and LLM-derived contextual representations, suggesting that frontal populations may preferentially encode sentence-level syntactic and contextual structure. A similar pattern was observed during an image question-answering task, suggesting that this tuning is not specific to sentence repetition alone. Notably, one array in middle frontal gyrus showed substantial overlap between motoric/phonemic and higher-level linguistic tuning, highlighting that very different levels of the speech hierarchy can overlap spatially within the same local neural population. These findings clarify the linguistic content of intracortical neural populations and identify sentence-level syntactic and contextual representations as promising targets for future neuroprosthetic decoding.

Topic Areas: Language Production,