A generalised model for utilising prosodic information in continuous speech recognition

Prosodic features in continuous speech provide cues which may be used to disambiguate syntactic ambiguities and to increase the accuracy of speech recognition/understanding systems. This paper presents a novel method using a multivariate statistical framework for producing a model of the relationship between prosodic and syntactic structures in continuous speech. The model can be used for Linguistic/Phonetic research and in speech synthesis. This paper concentrates on its use for integrating prosodic information into a continuous speech recognition system. The model can produce a relative probability score for the conformance of the prosodic and syntactic structures of hypothesised sentences for an existing word recognition system and achieves 73% accuracy in disambiguating structurally ambiguous sentences