Decoding dialogue is a matter of time

Abstract

Human conversation requires integration of language production and comprehension. However, the neural mechanisms at play during real-time conversations remain largely unexplored. A study by Yamashita and colleagues in Nature Human Behaviour has investigated the neural representation of conversational content across multiple timescales and across the brain. Participants lay in an MRI scanner engaging in conversations about 27 topics with the experimenter while functional MRI was recorded. The authors extracted contextual embeddings from conversation transcriptions — using a GPT model fine-tuned for interactive language tasks — and these were used to predict changes brain activity. Neural linguistic representations were partially shared between speech production and comprehension, but the topographic organization of these shared representations was modulated by timescales. Representations of words and single sentences (1–4 s long) were localized in association cortices, whereas those of longer contexts (16–32 s) were more unique and widespread. For language production, mean variance across regions was best explained by shorter context lengths, whereas longer contexts were best for language comprehension. These findings suggest that language production and comprehension have distinct temporal integration processes, reflecting the need for production to operate dynamically, whereas comprehension integrates linguistic input with broader contextual information.

Original reference: Nat. Hum. Behav. https://doi.org/10.1038/s41562-025-02231-4 (2025)