NLP models and neural dynamics of human conversation

• Natural language processing models have been used to characterize the neural dynamics of human conversation. ¹
• The work informs speech and language decoding and computational neuroscience. ¹
• It is of high relevance for communication BCIs; published in Nature (tier-1). ^{1 1}

Weekly enrichment (2026-07-20)

• The primary study is Cai, Hadjinicolaou, Paulk et al., “Natural language processing models reveal neural dynamics of human conversation,” published in Nature Communications (Nat Commun 16, 3376, 2025) by the Cash Lab at Massachusetts General Hospital/Harvard.²³
• Neural data came from semi-chronically implanted stereo-EEG (sEEG) depth electrodes in 14 intractable-epilepsy participants (6 female, 8 male; mean age 34, range 16–59) recorded during unconstrained, free-flowing conversation with an experimenter.²
• Recordings spanned 1,910 bipolar-referenced channels across 39 brain areas in both hemispheres; local field potentials were digitized at 2 kHz on Blackrock Cerebus systems and decomposed into alpha (8–13 Hz), beta (13–30 Hz), low gamma (30–55 Hz), mid gamma (70–110 Hz), and high gamma (130–170 Hz) envelopes.²
• Conversations lasted roughly one hour each (range 16–92 min) and yielded 86,637 transcribed words across participants, with 2,728 ± 1,804 words per participant during speech production.²
• Transcribed words were embedded with the GPT-2 language model (768 nodes per layer) and correlated channel-by-channel against LFP activity, using linear regression with Bonferroni correction across the 768 nodes to identify language-selective contacts.²
• Language-correlated activity was broadly distributed across frontotemporal cortex rather than localized, and neural patterns preferentially matched GPT-2’s middle and higher layers, implying encoding of contextual/sentence-level information rather than single words.²
• Contacts that tracked word sequences overlapped significantly with those marking conversational turn-taking: 39% (112/286) overlapped with speaker-to-listener transitions and 40% (84/210) of production-planning contacts changed at listener-to-speaker transitions, both far above chance (χ² p = 4.9 × 10⁻¹⁴ and 1.7 × 10⁻²²).²
• Established language regions retained a central role, with more than 20% of responding channels in superior temporal cortex during comprehension and left precentral cortex during production, while fewer than 20% of individual channels were shared between production and comprehension.²
• The authors frame semantic decoding — moving from identifying active regions to decoding word/concept meaning — as the next step, underscoring relevance to communication BCIs and speech neuroprostheses.³

Footnotes

1. https://news.google.com/rss/articles/CBMiX0FVX3lxTE44QVkycmotYkIwWGNXbVdSZ2IzQWRWWTd3cnFfMVNyWldoY2ZHc2NodUhmTDdGa2dxdURpeWhkSWJaRk5IVFc4TFg5RGdyUlp6ZjVYbldtNkYtRjNUaW9r?oc=5 ↩ ↩² ↩³ ↩⁴

2. https://www.nature.com/articles/s41467-025-58620-w ↩ ↩² ↩³ ↩⁴ ↩⁵ ↩⁶ ↩⁷ ↩⁸

3. https://www.massgeneralbrigham.org/en/about/newsroom/articles/using-ai-to-reveal-neural-dynamics-of-human-conversation ↩ ↩²