Note: Single-source report; awaiting corroboration.
A research team funded by the National Institutes of Health (NIH) used machine-learning models on single-cell brain recordings to identify neuronal activity associated with human speech production. Microelectrode arrays implanted in eight patients undergoing epilepsy monitoring recorded activity in the frontotemporal cortex during natural English conversations.
The researchers aligned neuronal data with transcriptions of the conversations and applied natural language processing models to examine the relationship between brain activity and language features. They discovered that neuronal signals immediately preceding speech could predict linguistic properties, including grammar, meaning, and sentence context, across diverse topics.
The study found a division of function among neurons: some encoded basic lexical information such as word meanings and grammatical roles, while others were linked to complex tasks like grouping words into structured sentences. Machine-learning models distinguished between similar phrases and words, suggesting that neuronal activity reflects unique contextual language elements.
According to the researchers, this level of cellular insight into speech offers a foundation for future studies and technologies to decode speech-related thoughts, potentially benefiting patients with communication impairments.