Improving posterior based confidence measures using enhanced local posteriors

In this paper, we propose a new technique for enhancing posterior probability based confidence measures in ASR systems. We propose to enhanced local posterior estimates used in confidence measurement process, in order to improve the overal confidence score in terms of ability to accept/reject a hypothesis. Posterior based confidence measures are global scores obtained by accumulating local evidences. These local evidences are usually phone posterior probabilities estimated in frame basis from speech signal. Having better (more informative) local evidences can potentially lead to better confidence measures. In [1, 2, 3], a method for enhancing local phone posterior estimates (evidences) has been proposed. This method is based on integrating prior knowledge (such as phone duration, lexical knowledge) and temporal context in the local posterior estimation. We show that using enhanced local posteriors in the confidence measurement process significantly and constantly improves their ability to predict whether a hypothesis (at word or phone level) is correct or incorrect, as compared to using regular local posterior estimates.