Conference on Cognitive Computational Neuroscience

Speech processing involves segmenting a continuous input stream at various levels (e.g., sounds or words). But does the brain also segment spoken words into their meaningful subparts (called morphemes)? We gathered neurophysiological (MEG) data from participants as they heard complex words in Arabic, and compared the data against three different models of speech comprehension: a naïve model without morphological features, a passive model with morpheme onset information, and a predictive model with boundary anticipation and morpheme surprisal and entropy. The predictive model explains significantly more data variability in bilateral superior temporal cortex compared to the passive model, which in turn explains more variability than the naïve model in bilateral temporal and inferior frontal regions. We also test different predictive parsing strategies. Our results support speech comprehension models that segment the input into morphemes predictively, rather than passively wait for boundaries or full morpheme identification.