Programming Spoken Dialogs Using Grammatical Inference

Abstract

Over-the-telephone Large Vocabulary Spoken Dialog Systems have now become a commercial reality. A major obstacle to the uptake of the technology is the effort required to construct spoken dialog applications, in particular the grammars. To overcome this obstacle, a spoken dialogue toolkit has been developed that uses grammatical inference in combination with a templating technique to build transaction based services. As part of this development a new grammatical inference technique know as the “Lyrebird” algorithm has been developed. Experimental results contained show that the Lyrebird algorithm outperforms the only other known algorithm for inferring context free attribute grammars. We also present the results of a comparison between the performance of the Lyrebird algorithm and an experienced speech application developer, showing that the algorithm creates grammars of a similar quality in a significantly reduced time.