This paper proposes a new technique to enhance speech understanding in spoken dialogue systems, which aims to replace semantic frames incorrectly generated by the systems with the correct ones. To do so, it relies on a training procedure that takes into account previous system misunderstandings for each dialogue state. Experiments have been carried out employing two systems (Saplen and Viajero) previously developed in our lab, which employ a prompt-independent language model and several prompt-dependent language models for ASR. The results show that the technique enhances system performance for both kinds of language model, especially for the prompt-independent language model. Using this technique, Saplen increases sentence understanding by 19.54%, task completion by 26.25% and word accuracy by 7.53%, whereas for Viajero these figures increase by 14.93%, 18.06% and 6.98%, respectively.

Index Terms: Spoken dialogue systems, speech recognition, speech understanding, dialogue management