Statistical Word Segmentation in Unfamiliar Speech

Statistical learning, the ability to detect patterns in sensory input, allows listeners to segment words from continuous speech by tracking transitional probabilities. While this mechanism is robust in familiar contexts, its adaptability to unfamiliar speech with distinct phonological properties remains less understood. This study investigates whether English-speaking adults can use TPs to segment an artificial language modeled on Cantonese. Participants identified words where syllables consistently occurred together (statistical words) and syllables that partially co-occurred (part-words) compared to those that never did (non-words). However, they struggled to distinguish statistical words from part-words when frequency was controlled. Pupillometry results showed participants dilated more to part-words and non-words at test, compared to frequency-controlled statistical words. Pupillary responses during familiarization also predicted test performance, demonstrating the potential of pupillometry to track learning in real time. These findings highlight the flexibility of statistical learning in adapting to novel linguistic contexts while revealing its limitations.