Linguistic processing in an island‐driven speech understanding system

Abstract

AbstractAn island‐driven control strategy in a speech understanding system (SUS) would be effective when reliable phonetic recognition cannot be obtained due to poor performance of the acoustic processor or noisy conditions. Especially, in conversational speech, the signal suffers from a heavy smoothing effect. An island‐driven SUS begins with detecting the acoustically most probable islands of keywords, and then expands them by a “hypothesis and test” paradigm until it covers the whole utterance using various sources of knowledge. However, this control strategy demands a very complicated linguistic processing rather than a left‐to‐right control. The authors developed an algorithm for transforming a set of syntactic rules written in the definite clause grammar into an efficient word predicting program in Prolog which operates under an island‐driven control. The basic ideas and the algorithm are explained here with a few examples of word prediction activities.