Robust speech understanding based on expected discourse plan

This paper reports spoken dialogue experiments for elderly people in the home health care system we have developed. In spoken dialogue systems, it is important to decrease recognition errors. The recognition errors, however, cannot be completely avoided with current speech recognition techniques. In this paper, we propose a robust recognition understanding technique based on expected discourse plans in order to improve a recognition accuracy. First, we collect dialogue examples of elderly users through a Wizard-of-Oz (WOZ) experiment. Next, we conduct a recognition experiment for collected elderly speech using the proposed technique. The experimental result demonstrates that this technique improved a sentence recognition rate from 69.1% to 74.3%, a word recognition rate from 80.3% to 81.7% , and a plan matching rate from 88.3% to 92.0%.