Automatic lexical generation for speech recognition

One essential step in the creation of the Sperry Univac word recognition systems is the production of a phonemic lexical representation for each word or phrase to be recognized. This document describes the automation of this lexical generation process. The procedure utilizes standard dictionary pronunciations as baseforms and produces lexical variants incorporating both phonological generalizations and systematic characteristics of the recognition system. The dictionary pronunciation guides are converted to system-compatible symbols and modified by rules which insert, delete, and replace segments. The rules are context sensitive, obligatory, and conjunctively ordered. The lexical strings produced reflect coarticulation effects both within and between words. Evaluation routines provide statistics on the use of the rules within a given lexicon. Initial application of the component shows word recognition results to be about 8% lower than those achieved with a hand-tuned lexicon.