Coping with speaker-related variation via abstract phonemic categories

Listeners can cope with considerable variation in the way that different speakers talk. We argue here that they can do so because of a process of phonological abstraction in the speech-recognition system. We review evidence that listeners adjust the bounds of phonemic categories after only very limited exposure to a deviant realisation of a given phoneme. This learning can be talker-specific and is stable over time; further, the learning generalizes to previously unheard words containing the deviant phoneme. Together these results suggest that the learning involves adjustment of prelexical phonemic representations which mediate between the speech signal and the mental lexicon during word recognition. We argue that such an abstraction process is inconsistent with claims made by some recent models of language processing that the mental lexicon consists solely of multiple detailed traces of acoustic episodes. Simulations with a purely episodic model without functional prelexical abstraction confirm that such a model cannot account for the evidence on lexical generalization of perceptual learning. We conclude that abstract phonemic categories form a necessary part of lexical access, and that the ability to store talker-specific knowledge about those categories provides listeners with the means to deal with cross-talker variation.