The perceptual generalization of normalized cue distributions across speakers

Listeners adapt to specific speakers' speech cue distributions and generalize the adaptation to the perception of a different speaker. It remains unclear whether listeners track and generalize the distributional statistics of raw, un-normalized cues or normalized cue distributions relative to the speaker's acoustic space. To address this question, we adopted a perceptual generalization paradigm to investigate whether manipulating context properties of a training speaker (Female A)'s speech would lead to different categorization results of critical phonemes in a test speaker (Female B)'s speech. Experiment 1 showed that learning Female A's speech containing the same set of sibilants but shifted vowel formants would lead to different categorization of Female B's sibilants: listeners exposed to raised vowel formants were more likely to identify an /s/ and those exposed to lowered vowel formants were more likely to identify /∫/ in Female B's speech, compared with participants exposed to unaltered vowel contexts. Experiment 2 showed that learning of Female A's speech containing the same set of stops but manipulated word frame duration would lead to different categorization results of Female B's stops: listeners temporally exposed to expanded word frames were more likely to identify a /t/ and those exposed to compressed contexts were less likely to identify a /t/ in Female B's speech, compared to participants exposed to unaltered temporal cues. These results suggest that listeners keep track of normalized cue distributions relative to the speaker's acoustic space and generalize those distributions to guide their speech perception behaviors. © 2022 The Author(s).