Estimated rank pruning and Java-based speech recognition

Abstract

Most speech recognition systems search through large finite state machines to find the most likely path, or hypothesis. Efficient search in these large spaces requires pruning of some hypotheses. Popular pruning techniques include probability pruning which keeps only hypotheses whose probability falls within a prescribed factor from the most likely one, and rank pruning which keeps only a prescribed number of the most probable hypotheses. Rank pruning provides better control over memory use and search complexity, but it requires sorting of the hypotheses, a time consuming task that may slow the recognition process. We propose a pruning technique which combines the advantages of probability and rank pruning. Its time complexity is similar to that of probability pruning and its search-space size, memory consumption, and recognition accuracy are comparable to those of rank pruning. We also describe a research-motivated Java-based speech recognition system that is being built at UCSD.