This paper presents EUROPA, a new generic framework for spoken dialogue systems, and its application to a car navigation task. A EUROPA-based system is based on keyword-spotting, i.e., it accepts and understands user's utterance as a set of keyword-sequences, to cope with the diversity of wording in spoken language. Applying voice interface techniques to practical tasks such as car navigation involves many recognition errors, false alarms in the case of keyword-spotting. EUROPA employs a new parsing algorithm, BTH, which is capable of efficiently parsing a keyword lattice that contains a large number of false alarms. The BTH parser runs without unfolding the given keyword lattice, and thus it can efficiently obtain a set of keyword-sequences acceptable to the given grammar as the parser result. This paper also presents MINOS, a prototype spoken dialogue system which is built by applying EUROPA to a car navigation task. It understands the user's spoken queries on either places or sections, and replies to them by synthesized voice. It runs on a single Windows-based portable PC in near real time.
展开▼
