Arthur Samuel
Department of Psychology
Stony Brook University

Perceptual learning for speech: Now you see it, now you don’t

There is rampant variation in speech, which means that the input a listener receives will differ from any canonical representation of a word that the listener might have. One way that people seem to cope with this variation is through perceptual learning: Hearing a speech sound that is not quite what the listener has in memory can cause a shift in the representation to make it more similar to the variant that was heard; if something like that variant is encountered again, the system should now match it better. Perceptual learning thus offers a potentially powerful way for the system to handle variation. However, shifting one’s representations has a possible cost in perceptual stability – it would be a bad idea to remap phonemic representations all the time. We have examined a wide range of conditions in which the system could potentially undergo perceptual learning, and we find that the system is impressively clever: It undergoes perceptual learning when the input variation seems to be characteristic of the speaker, but not if it can be attributed to some more transient factor. I will discuss a number of conditions in which perceptual learning does occur, and a number of conditions in which it does not. The pattern of shifts and of non-shifts indicates that the system is both nimble and stable.