Control concepts in a speech understanding system

Automatic speech understanding must accommodate the fact that an entirely accurate and precise acoustic transcription of speech is unattainable. By applying knowledge about the phonology, syntax, and semantics of a language and the constraints imposed by a task domain, much of the ambiguity in an attainable transcription can be resolved. This paper deals with how to control the application of such knowledge. A control framework is presented in which hypotheses about the meaning of an utterance are automatically formed and evaluated to arrive at an acceptable interpretation of the utterance. This design is currently undergoing computer implementation as a part of the Bolt Beranek and Newman (BBN) Speech Understanding System (SPEECHLIS).