Learning to understand an unfamiliar talker: Testing distributional learning as a model of rapid adaptive speech perception

Human speech perception is highly adaptive: exposure to an unfamiliar accent quickly reduces the difficulty listeners might initially experience. How such rapid adaptation unfolds incrementally remains largely unknown. This includes questions about how listeners' prior expectations based on lifelong experiences are integrated with the unfamiliar speech input, as well as questions about the speed and success of adaptation. We begin to address these knowledge gaps through a combination of an incremental exposure-test paradigm and model-guided data interpretation. We expose US English listeners to shifted phonetic distributions of word-initial "d" and "t" (e.g., "dill" vs. "till"), while incrementally assessing cumulative changes in listeners' perception. We use Bayesian mixed-effects psychometric models to characterize these changes, and compare listeners' behavior against both idealized learners (ideal observers that know the exposure statistics) and a model of adaptive speech perception (ideal adaptors that have to infer those statistics). We find that a distributional learning model provides a good qualitative and quantitative fit (R2>96%) to both listeners' prior perception and changes in their perception depending on the amount and type of exposure. We do, however, also identify previously unrecognized constraints on adaptivity that are unexpected under any existing model of adaptive speech perception: changes in listeners' perception seem to plateau below the level expected under successful learning.