Leveraging call context information to improve confidence classification

This paper describes how speech recognition confidence estimation in a typical Directory Assistance scenario can be improved by taking dialog context into account and recalibrating the original recognition confidences using a statistical classifier that employs classification features extracted from this context. We look at several types of classification features and investigate their utility with respect to semantic and sentence error rates with a view to an improved application behavior, but also with a long term goal of a more efficient semi-supervised selection of model training material. The method leads to significantly better tradeoffs between correct and false recognitions with respect to both error metrics.