Recent progress in automatic recognition of continuous speech

We have presented new search techniques for the second pass in our large-vocabulary continuous speech recognition system. In order to avoid the exponential growth of the tree structure in the depth first search, it is necessary to merge the recognition hypotheses in the course of the search. To further speed up the search, pruning is imposed at both word and phone levels where two thresholds are used in order to take into account the language model effect. We have also carried out branching ordering of the first pass word graph before doing the depth first search, which allows the envelope pruning to be more effective