What allows listeners to rapidly adapt to an unfamiliar talker? Evidence from a novel computationally-guided paradigm

Adaptivity in response to talkers with unexpected pronunciations is now understood to be central to speech perception. What mechanisms afford this ability, however, remains unclear. We present first results from a novel paradigm to investigate incremental changes in perception that occur with exposure to an unfamiliar talker. The paradigm is informed by a computational framework and provides high power to detect previously undocumented constraints on adaptive speech perception. L1-US English listeners (N = 122) were randomly assigned to one of three conditions, each corresponding to a different ‘accent’. Accents were identical except that the VOT distributions for /d/ and /t/ were shifted by either + 0msec, + 10msec or + 40msec relative to typical talkers of US English. We find that (1) listeners' categorization functions shifted proportionally with the input; (2) shifts were rapid, leading to significant changes after 48 observations; but (3) adaptation plateaued after 96 observations at only 40% of the shift in categorization that would have been justified by the statistics of the input. Result (2) and (3) together require revision of standard normalization, distributional learning, and exemplar/episodic accounts. We use computational models to show our results follow if listeners adapt by ‘mixing’ talker representations learned from previously experienced input.