Reducing peak search effort using two-tier pruning

Many of the pruning strategies used to remove less likely hypotheses from the search beam in large vocabulary speech recognition (LVR) systems, have a peak search space many times greater than the average search space. This paper discusses two such pruning strategies used within BT’s speech recognition architecture [1], Step pruning and Histogram pruning. Two-tier pruning is proposed as a simple but powerful extension applicable to either of the above strategies. This seeks to limit the expansion of the search space between the prune and acoustic match processes without affecting accuracy. It is shown that the application of two-tier pruning to either strategy reduces peak search effort, and results in an average reduction in run time of 33% and 53% for step pruning and histogram pruning respectively, with no loss in top-N accuracy.