Exploring phonotactic lexical constraints in word recognition

Abstract

In a previous meeting of the Society, Zue and Shipman demonstrated that the constraints imposed by the allowable sound sequences of a language are extremely powerful. Even at a broad phonetic level of representation, sequential constraints severely limit the number of possible word candidates [J. Acoust. Soc. Am. Suppl. 1 71, S7 (1982)]. This provides an attractive model for lexical access based on partial phonetic information. However, Zue and Shipman's results did not take into account the fact that the acoustic realizations of phonetic segments are highly variable, and this variability introduces a good deal of recognition ambiguity in the initial classification of the signal. We have conducted a set of studies investigating the robustness of sequential phonetic constraints with respect to variability and error in broad phonetic classification. In these studies segment misclassifications or deletions are permitted. In one study it was found that the phonetically variable parts of words (around reduced syllables) provide much less lexical constraint than the phonetically invariant parts. Thus, by utilizing the information from robust parts of a word, a large lexicon can still be partitioned into small equivalence classes. Detailed results of the studies will be presented. [Work supported by the Office of Naval Research under contract N00014-82-K-0727 and by the System Development Foundation.]