A two-pass decoding algorithm for partitioned search in continuous speech recognition

In this paper, a two-pass graph search technique is presented to conduct the search in the partitioned language graph. The first stage of the search uses a evaluation subroutine to efficiently reduce the graph to a subgraph consisting of the top “n” best paths. This small graph comprises a small search space of usually much less than size 𝒪(√N), whereas the original graph is or size N. The second stage of the search seeks out the optimal path from among those remaining in the subgraph. At the end of the search, the most likely path corresponding to the input utterance is found. This technique is applied to recognition of continuous speech taken from TIMIT speech database, and provides encouraging results