Novel two-pass search strategy using time-asynchronous shortest-first second-pass beam search

In this paper, we describe a novel two-pass search strategy for large vocabulary continuous speech recognition. The first-pass of this strategy uses a regular time-synchronous beam search with rough models to generate a word lattice. Then, the second-pass search derives exact results from the word lattice using more accurate models. This search is “time-asynchronous shortest-first beam search”, which has two novel features: a time-asynchronous beam search mechanism using heuristics that are scores on the word lattice nodes and a strict pruning scheme using shortest-first hypothesis extension. 20k-word Japanese broadcast news recognition experiments show that our second-pass search is more accurate and more efficient than either N-best rescoring or A* search that are conventional second-pass search methods.