In large vocabulary continuous speech recognition systems, one of the critical issues that have to be addressed is that of the search complexity. This paper studies two aspects of this problem. The first aspect is that of efficient pruning. We look at beam search and show that using a narrower beam for word endings can substantially improve the efficiency of the search. Secondly, we look at multipass search strategies and propose a first pass approach based on an approximate search using a tree representation of the lexicon. Experiments using this approach are presented.

Keywords: Large vocabulary speech recognition, beam search, lexicon tree, forward-backward search.