Conference on Cognitive Computational Neuroscience

How does language mediate brain-to-brain coupling during face-to-face communication between speakers and listeners? Here, we explore whether the embedding space learned by deep language models can serve as a common linguistic intermediary for aligning brains during communication. We recorded real-world, face-to-face conversations in five dyadic pairs of electrocorticography (ECoG) patients. This unique setup allows us to model how a speaker encodes and transmits their thoughts to a listener during free-form conversations. Our findings reveal the temporal profile of information flow across brains during conversation: linguistic content emerges in the speaker's brain before word onset and is recapitulated in the listener's brain rapidly following word onset. This transmission process relies on a shared linguistic embedding space for translating internal states from one brain to another.