An efficient beam pruning with a reward considering the potential to reach various words on a lexical tree

This paper presents an efficient frame-synchronous beam pruning for automatic speech recognition. With conventional beam pruning, hypotheses that have a greater potential to reach various words on a lexical tree are likely to be pruned out, since this potential is not taken into account. To make the beam pruning less restrictive for hypotheses with a greater potential and vice versa, the proposed method adds a reward as a monotonically increasing function of the number of reachable words from the node where a hypothesis stays on a lexical tree, to the likelihood of the hypothesis. The reward is designed not to collapse the ASR probabilistic framework. The proposed method reduces the processing time from 30% to 70% for grammar-based tasks. For a language-model-based dictation task, it also causes an additional reduction from the processing time of the beam pruning with the language model look-ahead technique.