Belief networks for a syntactic and semantic analysis of spoken utterances for speech understanding

In this paper we present a new approach towards speech understanding that merges semantic and intention decoding to one component. The algorithm is supposed to evaluate a speech recognizer’s utterance hypotheses regarding a) syntactical and semantical relations between words and phrases and b) potential intentions of the user. The mathematical fundament for this evaluation is probability theory. We make use of belief networks to handle the analysis of an utterance hypothesis as a process of reasoning with uncertain and incomplete information. The algorithm in general can be characterized as phrase spotting. The algorithm proved to be very robust for controlling the navigation system and the audio equipment of a car.