Third International Conference on Spoken Language Processing (ICSLP 94)

Yokohama, Japan
September 18-22, 1994

A Prosodic Recognition Module Based on Linear Discriminant Analysis

Andrew Hunt

Speech Technology Research Group, Dept. of Electrical Engineering, University of Sydney, NSW, Australia

Prosodic features in continuous speech are a potential source of useful information for speech recognition and understanding. This paper presents a technique for using knowledge of the relationship between acoustic prosodic features and syntactic structure to resolve syntactic ambiguity. The technique uses Linear Discriminant Analysis (LDA) to determine the linear combination of prosodic features which provides maximum separation of syntactic contexts and then applies a correction factor to compensate for topological variation. The LDA model determines the probability of conformance of prosody and syntax for rescoring candidate sentence hypotheses, and achieves 74% accuracy in resolving syntactic ambiguity on a standard corpus. The model is also of phonetic interest because it reveals a direct relationship between acoustic prosodic features and syntactic structure.

Full Paper

Bibliographic reference.  Hunt, Andrew (1994): "A prosodic recognition module based on linear discriminant analysis", In ICSLP-1994, 1119-1122.