An understanding strategy based on plausibility score in recognition history using CSR confidence measure

Although car-navigation systems attract attention as one of spoken dialogue interfaces, recognition errors due to the influence of natural speech and surrounding noise may prevent a smooth dialogue and disappoint the user. Thus, this research aims at the construction of a dialogue system which can achieve a smooth dialogue and a high degree of user satisfaction. Our system performs language understanding and response generation by using the confidence measure(CM) based on continuous speech recognizer(CSR) and the recognition history. This paper shows the spoken language understanding technique in the dialogue system. The CM, together with the speech type and the recognition history, is used for generating an integrated score. The system realizes a spoken language understanding which is more plausible for a given dialogue. As the result of evaluation experiment, it was shown that our system is more efficient (more than 15%) than a language understanding technique which simply gives priority to the higherrank hypothesis of a speech recognition result (n-best).