Proceedings of the Annual Meeting of the Cognitive Science Society

Deciphering the real-time transformation from acoustic signals to semantic meaning remains a challenge in cognitive neuroscience. Leveraging high-temporal-resolution magnetoencephalography data from native Chinese speakers and the Whisper computational framework, we investigated the neural mechanisms underlying speech processing. We demonstrated a robust scaling law, whereby larger models increasingly align with neural activity. Crucially, a model fine-tuned on native language experience outperformed generic multilingual models, underscoring the brain's sensitivity to language-specific statistics. Furthermore, variable-context analyses revealed a hierarchical temporal architecture: cortical integration of low-level acoustic features occurred within a rapid 40 ms window, whereas high-level speech representations required an integration window of approximately 400 ms. These findings provide a quantitative account of how native language experience and hierarchical temporal integration shape the neural encoding of human speech comprehension.