| Speech perception research often entails specific assumptions about the neural structures responsible for this complex behavior. Specifically, proponents of the Motor Theory of speech perception have posited an innate module with privileged access to auditory input. This module is assumed to map the incoming signal onto representations of articulatory gestures, bypassing the thorny problem of the lack of acoustic invariance in the constituent phonemes of a given utterance.;More recently, however, work on speech perception has framed the problem as a complex form of pattern recognition, in which words are directly mapped onto lexical representations, whether these are stored as groups of features, phonemes, or traces of previously encountered spoken words. On this view, speech perception is an expert behavior, and as such gives rise to special patterns of processing in laboratory situations.;In line with this more recent research, we first investigated auditory processing in the realm of second language learning. We trained neural network models to classify the phonetic inventory of English, then exposed them to novel speech contrasts from Zulu. The models' assimilation and discrimination behavior closely matched that of human participants. These results suggest that a general statistical learner is capable of acquiring a phonetic inventory in much the same way that people are.;In our second study, we trained participants to become expert perceivers of novel nonspeech stimuli. Participants completed a battery of psychophysical tests and were imaged (using fMRI) both before and after completing 20 hours of training on the novel sound distributions, in which they learned to map sound categories onto the blocks of the computer game Tetris. Behavioral and imaging results revealed long-term changes in the neural substrate of complex sound processing. Importantly, these changes correlated with participants' success in learning the training task. |
