Over the recent years, speech recognition technology has been making steady and significant progress. Together with advances in robust parsing techniques, natural language generation algorithms, and the advent of high-quality speech synthesis systems, it has paved the way for the emergence of robust spoken dialog systems. This paper has focused on the process involved in determining what the user has intended at each dialogue turn in a mixed initiative dialogue, conditioned on a recognizer word

more »

... graph with associated word confidence scores. The dialogue component directly influences the initial selection process at least whenever it has provided a specific context. The dialogue manager is the one component that has not only local information from each server, but also global knowledge about a particular user's constraints. In this paper, the Confidence Confirmation Algorithm in the selection of recognition hypotheses in the context of human-machine interactions is described. Enhancements also made to other human language technology servers for the purpose of providing useful information to the dialogue manager, as well as new capabilities in the dialogue manager itself aimed at detecting and repairing problematic spots in the dialogue.