Large vocabulary continuous speech recognition using N-best linear lexicon search and tree lexicon search with 1-best approximation

摘要

Computational cost on a large vocabulary continuous speech recognition system based on HMM is proportional to the number of words in the vocabulary. A tree-structured dictionary is generally used to reduce the number of states of HMMs. An approximation of dependency of word boundary and likelihood on word histories is also used to suppress the increase of hypotheses in the forward procedure. We first compared the search algorithms with a tree-structured dictionary using some approximation methods and that with a linear dictionary. The algorithm based on 1-best approximation with a tree-structured dictionary is efficient but frequently looses the optimal sentence hypothesis. Linear dictionary search can find the optimal hypothesis but needs much computational cost. Thus, we propose a search method using these two algorithms in parallel. We evaluated this new search algorithm and obtained improved word recognition rate and word accuracy by 5% and 3%, respectively on read speech, and 2% and 3%, respectively on broadcast news speech.