Conversational content is organized across multiple timescales in the brain

Masahiro Yamashita, Rieko Kubo, Shinji Nishimoto

Abstract: The evolution of conversation facilitates the exchange of intricate thoughts and emotions. The meaning is progressively constructed by integrating both produced and perceived speech into hierarchical linguistic structures across multiple timescales, including words, sentences and discourse. However, the neural mechanisms underlying these interactive sense-making processes remain largely unknown. Here we used functional magnetic resonance imaging to measure brain activity during hours of spontaneous conversations, modelling neural representations of conversational content using contextual embeddings derived from a large language model (GPT) at varying timescales. Our results reveal that linguistic representations are both shared and distinct between production and comprehension, distributed across various functional networks. Shared representations, predominantly localized within language-selective regions, were consistently observed at shorter timescales, corresponding to words and single sentences. By contrast, modality-specific representations exhibited opposing timescale selectivity: shorter for production and longer for comprehension, suggesting that distinct mechanisms are involved in contextual integration. These findings suggest that conversational meaning emerges from the interplay between shared linguistic codes and modality-specific temporal integration, facilitating context-dependent comprehension and adaptive speech production. Yamashita et al. explore how conversational content is represented in the brain, revealing shared and distinct brain activity patterns for speech production and comprehension, with contrasting timescale properties between the two processes.

External IDs:doi:10.1038/s41562-025-02231-4