Abstract: In this thesis we tackle the problem of identifying the best practices when designing and evaluating a spoken dialogue system. With the purpose of demonstrating that a more natural, flexible and robust dialogue is possible, and introducing a spoken dialogue system for controlling a Hi-Fi audio system as the selected prototype, we propose a Bayesian Networks (BNs) based solution for dialogue modelling combined with carefully designed contextual information handling strategies. Dynamic capabilities are also provided to keep the dialogue context permanently updated according to the evolution of the dialogue. All the thesis contributions have been evaluated finding an experimental support enough to demonstrate their relevance.
Index Terms: spoken dialogue systems, mixed initiative, Bayesian Networks, contextual information, usability, real users, evaluation, electronic devices control.