Bears don't always mess with beers: Limits on generalization of statistical learning in speech.

Perception changes rapidly and implicitly as a function of passive exposure to speech that samples different acoustic distributions. Past research has shown that this statistical learning generalizes across talkers and, to some extent, new items, but these studies involved listeners' active engagement in processing statistics-bearing stimuli. In this study, we manipulated the relationship between voice onset time (VOT) and fundamental frequency (F0) to establish distributional regularities either aligned with American English or reversed to create a subtle foreign accent. We then tested whether statistical learning across passive exposure to these distributions generalized to new items never experienced in the accent. Experiment 1 showed statistical learning across passive exposure but no generalization of learning when exposure and test items shared the same initial consonant but differed in vowels (bear/pear → beer/pier) or when they differed in initial consonant but shared distributional regularities across VOT and F0 dimensions (deer/tear → beer/pier). Experiment 2 showed generalization to stimuli that shared the statistics-bearing phoneme (bear/pear → beer/pier), but only when the response set included tokens from both exposure and generalization stimuli. Moreover, statistical learning transferred to influence the subtle acoustics of listeners' own speech productions but did not generalize to influence productions of stimuli not heard in the accent. In sum, passive exposure is thus sufficient to support statistical learning and its generalization, but task demands modulate this dynamic. Moreover, production does not simply mirror perception: generalization in perception was not accompanied by transfer to production.