Decoding network optimization using minimum transition error training

The discriminative optimization of decoding networks is important for minimizing speech recognition error. Recently, several methods have been reported that optimize decoding networks by extending weighted finite state transducer (WFST)-based decoding processes to a linear classification process. In this paper, we model decoding processes by using conditional random fields (CRFs). Since the maximum mutual information (MMI) training technique is straightforwardly applicable for CRF training, several sophisticated training methods proposed as the variants of MMI can be incorporated in our decoding network optimization. This paper adapts the boosted MMI and the differenced MMI methods for decoding network optimization so that state transition errors are minimized in WFST decoding. We evaluated the proposed methods by conducting large-vocabulary continuous speech recognition experiments. We confirmed that the CRF-based framework and transition error minimization are efficient for improving the accuracy of automatic speech recognizers.

1. INTRODUCTION

Current speech recognizers involve several probabilistic models, such as acoustic models, lexicons, and language models. Therefore, the efficient integration of these models is important if we are to optimize the speech recognition performance. Conventionally, these probabilistic models are trained separately, and integrated simply by taking the product of probabilities. This basic scheme is consistent if we consider the maximum likelihood (ML) training of hidden Markov model (HMM)-based acoustic models, and N-gram language models. This is because these models are designed so that the joint objective function can be factorized into a likelihood term and a prior probability term computed by acoustic and language models, respectively.