Revisiting Distinctive Feature Approach In Speech Recognition

Abstract

This paper intends to show how current phonological theories of feature hierarchy and underspecification can be implemented into speech recognition so that we can take advantage of both invariant and allophonic cues for parsing and matching. I first show the usefulness of variant cues in parsing without relying on the higher level constraints, and I show that the distinctive feature approaches are better than the segmental approaches in confusion error analyses as well as in utilizing phonological generalizations. Mainly due to the defects of the feature theory per se, the feature analysis has been neglected in speech recognition, and thus I argue that the binary notion of feature values and the requirement of fully specified phonetic representation should be given up, and instead the privativeness notion of features be adopted, which leads to surface phonetic underspecification.