Phonetics–phonology mapping in the generalization of perceptual learning

Previous studies on whether perceptual learning generalizes across multiple speakers have produced inconsistent results between generalization and speaker-specificity. A prior proposal is that the critical phonemes produced by two different speakers need to be phonetically similar for perceptual learning to generalize. To test this account, we investigated the perceptual generalization of sibilants across two pairs of speakers. In both cases, sibilants of the same male speaker were manipulated to induce either an /s/-favoring perceptual bias or an /ʃ/-favoring one in a training phase. We then examined whether the perceptual biases would generalize in a test phase to the /s/-/ʃ/ continua of two different female speakers, one resembling the training speaker and the other differing from the training speaker in the spectral frequency distributions of their sibilants. We found that generalization of perceptual learning occurred in both /s/-favoring and /ʃ/-favoring conditions between the speaker pair with similar sibilant productions. For the speaker pair with different sibilants, we found perceptual generalization in the /s/-favoring training condition but not in the /ʃ/-favoring condition, which is not predicted by the phonetic similarity account. To explain these unexpected results, we offer a novel phonetics–phonology mismatch account as a refinement of our understanding of when and why perceptual generalization might be blocked. The results shed light on the constant influence of the mapping between phonetics and phonology during the learning and generalization of phonetic variability.