Knowledge-based systems in speech recognition: a survey

Abstract

Endeavours in large vocabulary (1000 word) automatic speech recognition (ASR) systems for task-oriented dialogues in which the syntax, semantics and pragmatics are inherently limited enjoy success within their domain, but it may prove impractical to build larger systems based on the techniques which enable this.

There is a consensus of opinion among researchers in ASR that the application of speech-specific knowledge would facilitate the design of more sophisticated stochastic models than those being attempted currently, and reduce the amount of training data necessary for an acceptable level of performance. There have been several attempts at designing purely knowledge-based ASR systems and, more recently, hybrid classifiers which combine stochastic techniques and speech knowledge: some of these systems are described in this survey. Whichever approach is adopted, there are common problems of knowledge representation and control, which may continue to impede the progress of ASR until knowledge-based techniques progress.