Recognition Confidence Scoring for Use in Speech Understanding Systems

In this paper we present an approach to recognition confidenc e scoring and a method for integrating confidence scores into t he understanding and dialogue components of a speech understa nding system. The system uses a multi-tiered approach where co nfidence scores are computed at the phonetic, word, and uttera nce levels. The scores are produced by extracting confidence fea tures from the computation of the recognition hypotheses and proc essing these features using an accept/reject classifier for word an d utterance hypotheses. The output of the confidence classifiers can then be incorporated into the parsing mechanism of the language u nderstanding component. To evaluate the system, experiment s were conducted using theJUPITERweather information system. Evaluation was performed at the understanding level using key-v alue pair concept error rate as the evaluation metric. When confid e ce scores were integrated into the understanding component of the system, the concept error rate was reduced by over 35%.