Yossi Matias

This study shows that speakers' brains generate information-rich words through extra neural computation, evidenced by stronger neural activity before articulation, and that LLMs similarly require deeper computation for improbable words.

The paper uses a deep speech-to-text model to predict neural responses during spontaneous conversations, revealing a distributed cortical hierarchy for speech and language processing.

In nine epilepsy patients, cortical activity carries next-word information up to 800 ms before onset, tracks GPT-2 surprise after onset, and fits GPT-2 contextual embeddings better than static ones.