An efficient layer-wised beam pruning algorithm for large vocabulary continuous speech recognition system

This paper presents a layer-wised pruning algorithm for time-synchronous search in large vocabulary continuous speech recognition system. Based on the traditional beam pruning and histogram pruning algorithm, we propose a simple but efficient pruning strategy by taking the depth of hypothesis in decode network into calculation of its beam width. The depth of hypothesis could be used for evaluating its uncertainty. The hypothesis with more uncertainty will use a more conservative and looser beam width than the hypothesis with lower uncertainty. Experiments in various tasks show that layer-wised pruning can accelerate our highly optimized decoder by 10%-30% relatively without impacting the precision.