Selecting, Designing, and Using Practical Speech Recognizers

Abstract

An expanding public awareness of the value of voice command of machines has made it critical that the previous trial-and-error methods for buying, using, and designing speech recognizers be replaced by systematic procedures. A methodology has been developed with Honeywell for identifying frequently-occurring discrete tasks in high-workload situations for which manual machine control is difficult and the utility of voice input is high. The feasibilities of voice in such tasks have also been assessed, based on the confusabilities of the vocabularies, the noise conditions, and other factors. The detailed specifications of current devices that permit recognition of isolated words and restricted word sequences have been compiled, and criteria defined for selecting the best device for each purpose. Recently, initial side-by-side comparisons of recognizers have been made, and tests have been designed for systematic testing of devices under a variety of representative conditions. The design of better recognizers should begin with consideration of features that assure competitiveness, special facilities that aid the user, and algorithms that are not critically dependent upon word boundaries, equal attention on all acoustic data, or simplistic template matching. It is now possible to advance the state of the “art” of speech recognition to more of a systematic “science” for effective buying, trying, applying, and modifying of speech recognizers.