Beam search pruning in speech recognition using a posterior probability-based confidence measure

In this work we propose the early incorporation of confidence information in the decoding process of large vocabulary speech recognition. A confidence based pruning technique is used to guide the search to the most promising paths. We introduce a posterior probability-based confidence measure that can be estimated efficiently and synchronously from the available information during the search process. The accuracy of this measure is enhanced using a discriminative training technique whose objective is to maximize the discrimination between the correct and incorrect decoding hypotheses. For this purpose, phone-level confidence scores are combined to derive word level scores. Highly compact models that exhibit minimal degradation in performance are introduced. Experimental results using large speech corpora show that the proposed method improves both the decoding accuracy and the decoding time when compared to a baseline recognition system that uses a conventional search approach. Furthermore, the introduced confidence measures are well-suited for cross-task portability.