Despite great progress in natural language understanding and generation, we still know very little about how our brains do these tasks. More insight on human language processing could lead to smarter AIs that can be trained with less data. Our project focuses on determining how humans extract linguistic features from spoken language. Using a rich dataset of stereoelectroencephalography (sEEG) recordings in response to audiovisual stimuli, we can analyze brain activity while various phrases are spoken with high spatiotemporal resolution. Recent work in this field already shows promise, with ECoG being sufficient to produce speech sounds. We hope to take this work one step further by finding correlates in the brain for underlying phonetic feature of language, but a prerequisite to this is understanding how speaker identification is done in the brain. Different speakers can say the same words in drastically different ways, but the brain somehow normalizes against this. Our current work focuses on this speaker identification and normalization process in the brain. Preliminary results have identified various regions, previously suspected to be integrating auditory and visual information, that have strong speaker-related signals as soon as 300ms post-stimulus onset.

“Through this SuperUROP, I hope to contribute meaningfully in a field I’m very passionate about, as well as learn more about the research culture. After many machine learning, neuroscience, and linguistics courses, I have seen how difficult it is to model temporal dependencies, and how necessary these dependencies are for language. I hope to use my skills to deepen our understanding of the brain’s processing of language, and more broadly, time.”