HMN 2026: How With neuronal data, AI models predict grammar, meaning and context of spoken sentences

By applying machine-learning models to single-cell brain recordings taken from humans in conversation, a research team identified both individual and collective neuronal activity that reflected key features of language. The work, published in Nature, offers unprecedented insight into how neurons encode linguistic information, suggesting that brain activity may one day be used to infer speech-related thoughts, which could be transformative for some patients.

“This level of granularity is necessary for us to more completely understand how the brain generates speech and, ultimately, how we can develop technologies to restore it for individuals with communication disorders,” said Debara Tucci, M.D., director of NIH’s National Institute on Deafness and Other Communication Disorders (NIDCD).

Conversations offer a rare window

The neuronal data came from microelectrode arrays implanted in eight patients for the separate purpose of epilepsy monitoring. The scientists, from Massachusetts General Hospital in Boston, made use of the opportunity by conducting and recording naturally flowing conversations in English, spanning a wide range of topics, with each of the study participants.

The researchers aligned transcriptions of the conversations in time with data describing the activity of hundreds of neurons in the frontotemporal cortex—a region the team previously linked to speech production. Then, using natural language processing models, they set out to uncover relationships between the data sets.

Neurons split language tasks

The authors found that neuronal recordings from just before participants spoke predicted many properties of subsequent speech, across any topic of discussion. They detected a division of labor among the examined neurons, with some reflecting basic information, such as the meaning and roles of specific words, while others tackled more complex tasks, including grouping phrases into structured sentences.

Their models could distinguish between similar phrases and words, suggesting the neuronal activity captured the unique context of sentences as well.

“For the first time, we’re describing processes not only at the regional but at a cellular scale that produce speech. Having identified these fundamental building blocks, we’ve set the table for us to begin answering some really interesting questions,” said first author Jing Cai, Ph.D., a researcher and instructor at Mass General.

These findings reveal how individual neurons encode language during speech, advancing understanding of the brain’s linguistic architecture. This knowledge could enable a new generation of technologies that translate neural activity into machine-generated speech beyond current capabilities.

The content is provided for information purposes only.