Idiosyncrasy and generalization in accent learning

People understand speech well, despite pronunciation variation. Perceptual learning, where listeners are trained with acoustic features ambiguous between two phonemes and subsequently shift their perceived phoneme boundary, is one way listeners may compensate for variation (Norris et al 2003). These perceptual shifts, however, seem idiosyncratic to one speaker (Eisner & McQueen 2005, Kraljic & Samuel 2005), rarely generalizing to new speakers. We propose that lack of generalization is due to lack of experience mapping phonemes to specific continua; previous work uses continua like [s]-[f], whose midpoints rarely occur in speech. Ambiguous tokens that are never heard in real speech may be perceived as specific to the speaker used in training, preventing generalization. Use of a continuum occurring in accented speech, such as the mapping of English tenseness onto vowel duration, allows manipulation of the idiosyncrasy of the mapping. We train 13 listeners on an idiosyncratic duration mapping (lax to short d...