Date of Award


Publication Type

Doctoral Thesis

Degree Name



Computer Science

First Advisor

Richard A. Frost (Computer Science)


Artificial Intelligence.




Currently, most commercial speech-enabled products are constructed using grammar-based technology. Grammar design is a critical issue for good recognition accuracy. Two methods are commonly used for creating grammars: 1) to generate them automatically from a large corpus of input data which is very costly to acquire, or 2) to construct them using an iterative process involving manual design, followed by testing with end-user speech input. This is a time-consuming and very expensive process requiring expert knowledge of language design, as well as the application area. Another hurdle to the creation and use of speech-enabled applications is that expertise is also required to integrate the speech capability with the application code and to deploy the application for wide-scale use. An alternative approach, which we propose, is 1) to construct them using the iterative process described above, but to replace end-user testing by analysis of the recognition grammars using a set of grammar metrics which have been shown to be good indicators of recognition accuracy, 2) to improve recognition accuracy in the design process by encoding semantic constraints in the syntax rules of the grammar, 3) to augment the above process by generating recognition grammars automatically from specifications of the application, and 4) to use tools for creating speech-enabled applications together with an architecture for their deployment which enables expert users, as well as users who do not have expertise in language processing, to easily build speech applications and add them to the web.