Incorporating a feature tree geometry into a matcher for a speech recognizer

Abstract

The goal of this thesis is to incorporate a feature tree geometry into an existing matcher for a lexical access system. Proposed by Professors Kenneth N. Stevens and Samuel Jay Keyser in their paper Feature Geometry and the Vocal Tract, the feature tree geometry is a hierarchical structure for representing the features of a phoneme. The four components involved in incorporating the feature tree geometry into the lexical access system are examined individually: the development of a computational model, the invocation of rules to account for assimilation in running speech, the production of morphemic stems, in particular, plurals, and the process of word retrieval. The last component is not implemented with the feature tree, but the implications of a tree-based matcher are explored. Initially, a brief overview of the lexical access system is presented for the reader's background. Each component is discussed first from a theoretical standpoint and then its implementation is presented and discussed. This paper concludes with a comparison of the original matcher's performance to the performance of the matcher based on the feature tree geometry. Performance results indicate that the computation time of the matcher fitted with the feature tree geometry decreased to 40% of the original matcher's execution time. Overall, this paper represents an initial probe into using alternative lexical representations in a lexical access system to obtain a desired level of performance.