Research recently published in journals nature We provide evidence that individual brain cells in the human cortex function as specialized building blocks for constructing the complex structure of spoken language. By recording electrical activity directly from the brains of people engaged in natural conversation, researchers discovered that specific groups of neurons are dedicated to processing different parts of speech, sentence structure, and meaning. These findings suggest a detailed cellular map of how we produce speech, providing insights that may ultimately help develop advanced techniques to restore communication in individuals with language disorders.
Language is a distinctly human ability that allows us to communicate an infinite variety of thoughts. “Language is one of the key characteristics of being human, but surprisingly little is known about how individual brain cells make language possible,” says study author Jing Tsai, principal investigator at the China Brain Research Institute. At a broad level, brain imaging scans show that a network of regions throughout the frontotemporal cortex is at work when we speak.
The frontotemporal cortex is a region located near the front and sides of the brain that plays a key role in language and cognition. Neuroimaging highlights general areas of brain activity but does not reveal the microscopic cellular processes involved in natural speech. Although brain scans show the big picture, they cannot show how individual brain cells, known as neurons, encode grammatical categories or relationships between words.
“Human language is one of the most remarkable aspects of human cognition,” Cai says. “It’s hierarchical, highly structured, and we can easily communicate with each other. My background in machine learning and large-scale language models naturally led me to ask how individual neurons in the human brain process language, and whether the representations that emerge at the single-neuron level share any similarities with the representations found in LLM.”
Large-scale language models (LLMs) are advanced artificial intelligence programs trained on vast amounts of text to recognize and predict human language patterns. Examining brain activity at this microscopic level can help reveal how linguistic information is distributed in different brain regions. Scientists want to know whether single neurons recognize the overall structure of a sentence and how they encode different phrases.
To map these cellular processes, the authors recorded brain activity in eight human participants, including three women and five men, with an average age of 40 years. These were patients who were already scheduled to undergo surgical monitoring for severe epilepsy. As part of their medical care, tiny microelectrode arrays were temporarily implanted in their brains. These small devices contain a grid of sensors that can detect the electrical signals, or action potentials, of individual neurons.
The scientists tracked the activity of 579 individual neurons in isolation over 14 separate sessions. During the recording, participants engaged in natural, unscripted conversation. They answered a variety of questions and prompts and discussed topics ranging from personal feelings and spatial location to health and opinions. In total, participants produced 10,460 words in 1,895 independently constructed sentences.
Scientists synchronized audio recordings of these conversations with electrical activity recorded from neurons. They used natural language processing models to analyze spoken texts. These programs work like automatic grammar teachers, labeling each word based on its part of speech, grammatical role, and position within the broader structure of the sentence.
Two different types of text analysis were used to classify words. Constituency analysis breaks down sentences into nested structural units, such as noun phrases and verb phrases. Dependency analysis maps direct grammatical links between specific words, such as adjectives that modify nouns. The researchers looked for connections between these mathematical descriptions and the cells’ firing patterns.
The data revealed a highly specialized division of labor between neurons. Approximately 9 percent of the recorded cells responded preferentially to certain parts of speech. These neurons increased their electrical firing just before participants said certain types of words, such as nouns or verbs.
“We were surprised by the amount of information that individual neurons carried,” Cai told PsyPost. “Some neurons encoded detailed grammatical relationships, while others tracked higher-order sentence structure and meaning.”
For example, about 16% of the neurons monitored the depth of the word hierarchy. This means tracking how deeply words are embedded within the grammatical branches of a sentence. An additional 10% of cells tracked dependencies, changing their activity based on whether the planned word served as a direct object or subject.
The researchers also found that individual neurons tend to separate meaning from grammar. Most language response cells specialize in encoding either sentence structure rules or word definitions, but rarely both. Only about 2% of the recorded cells encoded both syntactic and semantic information simultaneously.
“With the help of LLM, our findings show that single neurons do not simply respond to individual words, but work together to represent grammar, meaning, and sentence structure in a flexible and combinatorial way,” Cai said. As a group, the cells accurately captured a combination of grammatical and semantic features of speech. This suggests that the brain uses a distributed network of specialized cells to build complete representations of language.
To investigate how these cells process the broader context of speech, the scientists used large-scale language models to map out how a word’s meaning changes based on the words that come before it. The authors found that the neurons dynamically adjusted their firing patterns based on the context of the sentence, successfully incorporating information from up to five preceding words.
“More surprisingly, these neurons dynamically adapt their responses depending on the context of the sentence, suggesting that single cells are involved in the highly flexible representation of language,” Cai said. This anticipatory brain activity peaked about a second before the participants actually spoke the words.
To confirm that the neurons were indeed responding to linguistic context, the researchers performed controlled tests using randomly swapped sentences and meaningless alternative words. Once the model processed these scrambled inputs, it could no longer predict the neurons’ firing patterns. This shows that brain cells were actively tracking the true meaning and flow of the conversation.
The distribution of these specialized neurons has also provided new insights into the organization of the brain. It was found that language-responsive cells are widely scattered in the frontal and temporal lobes. However, the strength of their responses was not the same everywhere. Neurons located in the left hemisphere of the brain responded significantly more strongly to linguistic features than neurons located in the right hemisphere.
This is consistent with the common medical understanding that the left hemisphere tends to dominate language tasks in most people. The authors also compared the activity of individual neurons to the general background electricity in the surrounding brain tissue, known as the local field potential. Local field potentials measure the synchronized activity of thousands of cells in close proximity.
The researchers found that individual neurons were much more precise and specialized in language regulation than broad brain waves measured in the exact same location. Although a particular microscopic region may have a common electrical response to parts of speech, individual neurons located at that precise region are often tuned to completely different linguistic features. Individual cells appear to act as highly specialized filters, even when neighboring cells are engaged in other tasks.
Although this study provides a detailed look at the cellular basis of language, it does have limitations. “This is an early map of how individual neurons encode language, not the whole picture,” Cai explained. “Future studies should record from more brain regions, study other forms of communication such as language comprehension and writing, and test whether these findings generalize to different situations and populations.”
The analysis also does not address how neurons encode expressive elements of speech, such as tone, pitch, and emotional intonation. Because this study included participants with epilepsy, it is also possible that underlying neurological conditions may influence some aspects of brain activity. However, the researchers specifically selected brain regions where language function was intact to minimize this risk.
As researchers continue to map these cellular components, their findings could support the development of new medical technologies. “This brings us closer to understanding how the brain generates language and provides the foundation for the development of future brain-computer interfaces that can restore communication to people who have lost the ability to speak,” Cai said.
The study, “Mapping the neural components of human language with language models,” was authored by Jing Cai, Yoav Kfir, Mohsen Jamali, Hesen Huang, Young Joon Kim, Sydney S. Cash, and Ziv M. Williams.
