Contributed Talk Sessions | Poster Sessions | All Posters | Search Papers

Poster Session B: Wednesday, August 13, 1:00 – 4:00 pm, de Brug & E‑Hall

Intracranial EEG reveals multiplexed encoding of auditory, speech, and language embeddings in the human temporal lobe

Atlas Kazemian¹, Josef Parvizi¹, Daniel LK Yamins¹, Laura Gwilliams¹; ¹Stanford University

Presenter: Atlas Kazemian

Speech understanding in the human brain involves several representational transformations: Air pressure fluctuations become a time-frequency representation in the cochlear; auditory cortex extracts speech-relevant units; distributed networks extract amodal representations of meaning and structure. Recent advances in speech and language models have led to a series of studies using text-based large language model (LLM) representations to model these neural transformations. Here we examine the brain’s end-to-end processing of language by extending the investigation to include biophysical models of the cochlear and auditory cortex, as well as performance-optimized models of speech (Whisper) and text (GPT-2, Llama-3). We use these models to predict the activity of spatially precise neural populations recorded via intracranial EEG (iEEG) as participants listened to audiobooks. Our findings are twofold. First, we observe clear differences in encoding performance within both auditory and language model families. Second, each model type captures distinct aspects of the signal in temporal lobe electrodes, suggesting that these regions encode a mixture of intermediate auditory and higher-level semantic features. Together, these results highlight the importance of examining model–brain alignment with fine-grained temporal precision.

Topic Area: Language & Communication

Extended Abstract: Full Text PDF