A unified theory of composite pattern analysis for automatic speech recognition.

Words are combined both probabilistically and deterministically to form dialogs at the semantic level. Two parallel mathematical descriptions, parametric and nonparametric, are available at each level. A description at one level is made applicable to its successor simply by generalizing the appropriate mathematical concepts in a natural way. The resulting unified formulation has the virtue of being directly translatable into a simple elegant computer program which is capable of conducting a circumscribed but non-trivial dialog with a human. The architecture of the program is described and a demonstration of its performance is given.