Dialogue Strategies to Overcome Speech Recognition Errors in Form-Filling Dialogue

摘要

In a spoken dialogue system, the speech recognition performance accounts for the largest part of the overall system performance. Yet spontaneous speech recognition has an unstable performance. The proposed post processing method solves this problem. The state of a legacy DB can be used as an important factor for recognizing a user's intention because form-filling dialogues tend to depend on the legacy DB. Our system uses the legacy DB and ASR result to infer the user's intention, and the validity of the current user's intention is verified using the inferred user's intention. With a plan-based dialogue model, the proposed system corrected 27% of the incomplete tasks, and achieved an 89% overall task completion rate.