Sequence-Discriminative Training of Neural Networks

In this chapter we explore sequence-discriminative training techniques for neural-network–hidden-Markov-model (NN-HMM) hybrid speech recognition systems. We first review different sequence-discriminative training criteria for NN-HMM hybrid systems, including maximum mutual information (MMI), boosted, minimum phone error, and state-level minimum Bayes risk (sMBR). We then focus on the sMBR criterion, and demonstrate a few heuristics, such as denominator language model order and frame-smoothing, that may improve the recognition performance. We further propose a two-forward-pass procedure to speed up sequence-discriminative training when memory is the main constraint. Experiments were conducted on the AMI meeting corpus.