EFFICIENT SEARCH USING POSTERIOR PHONE PROBABILITY ESTIMA TES

In this paper we present a novel, efficient search strategy for large vocabulary continuous speech recognition (LVCSR). The search algorithm, based on stack decoding, uses posterior phone probabil- ity estimates to substantially increase its efficiency with minimal effect on accuracy. In particular, the search space is drama tically reduced by phone deactivation pruningwhere phones with a small local posterior probability are deactivated. This approac h is par- ticularly well-suited to hybrid connectionist/hidden Mar kov model systems because posterior phone probabilities are directl y com- puted by the acoustic model. On large vocabulary tasks, using a trigram language model, this increased the search speed by an order of magnitude, with 2% or less relative search error. Re sults from a hybrid system are presented using the Wall Street Journal LVCSR database for a 20,000 word task using a backed-off trigram language model. For this task, our single-pass decoder took around 15× realtime on an HP735 workstation. At the cost of 7% relative search error, decoding time can be speeded up to approximately realtime.