Temporal structure of the language hierarchy within small cortical patches

A neuroscience team from institutions including École Normale Supérieure has published a groundbreaking study on the neural basis of speech production. By analyzing ultra-high-resolution recordings from eight 64-microelectrode arrays implanted in the motor cortex and inferior frontal gyrus of two patients, the researchers captured brain activity during the production of 20,000 sentences. Their key finding challenges traditional models of macroscopic brain organization for language. Instead of finding distinct brain regions dedicated to specific linguistic levels (like phonemes or words), they discovered that small cortical patches—just 3.2x3.2 mm in size—robustly encode a multiplexed hierarchy of linguistic features.

Within these tiny patches, the brain simultaneously represents phonetic, syllabic, and lexical information. Critically, this coding scheme is not static but dynamically changes over time. This temporal multiplexing allows successive phonemes, syllables, and words to be represented without neural interference, enabling the rapid coordination required for fluent speech. The authors explicitly note that this neural architecture is reminiscent of the position encoding mechanism used in transformer-based AI models like GPT-4, where information about sequence order is dynamically integrated. This provides a concrete biological parallel to artificial neural networks, suggesting convergent computational strategies for handling hierarchical, sequential information.