Implementing and Improving MMIE Training in SphinxTrain

Discriminative training schemes, such as Maximum Mutual Information Estimation (MMIE), have been used to improve the accuracy of speech recognition systems trained using Maximum Likelihood Estimation (MLE). In this paper, we present the implementation details of MMIE training in SphinxTrain and baseline results for MMIE training on the Wall Street Journal (WSJ) SI84 and SI284 data sets. This paper also introduces an efficient lattice pruning technique that both speeds up the process and increases the impact of MMIE training on recognition accuracy. The proposed pruning technique, based on posterior probability pruning, is shown to provide better performance than MMIE using standard pruning techniques.