Against Underspecification in Speech Errors

This paper argues against the use of phonological underspecification in feature matrices on the basis of speech error data. Stemberger 1991 argues that phonological underspecification influences the similarity of phonemes. He claims underspecified features do not count toward similarity, based on an analysis of phoneme confusions in a naturally occurring speech error corpus. Using the same corpus, I show that a similarity metric that does not employ underspecification provides a better account of the data. This similarity metric, originally proposed in Frisch, Broe, & Pierrehumbert 1995, is sensitive to featural redundancy but does not omit redundant features. In this metric, redundant features contribute less toward similarity than contrastive features. These results show that redundant features play a role in predicting speech errors, and thus that phonological representations must encode redundancy and not exclude it.