Research into human speech recognition reveals an apparent conflict between the high resolution of the speech processing mechanisms and the variability of the speech input. Access to the lexicon is disrupted even by single feature mismatches, while fluent speech is marked by a wide variety of phonological processes which change the surface form of the words being uttered. The word sweet, for example, may be produced as [swit], [swik] or [swip], depending on its right context. This is a problem for a theory of lexical access, since sweek is a nonword, while sweep is an existing word in the language, and yet in each case the correct outcome is the recognition that the word sweet is intended. In a series of studies of this phenomenon, using gating and priming techniques, we find that phonologically regular distortions of the speech input do not create mismatch in lexical access, so long as these distortions occur in the appropriate phonological context. We discuss the implications of these findings for a theory of the speech recognition process, emphasising the abstractness of lexical form representations and the on-line involvement of processes of phonological inference in determining the correct interpretation of the incoming speech signal.

Keywords: Phonology, Variation, Lexical Access, Mental Representations.