An efficient lexical tree search for large vocabulary continuous speech recognition

This paper describes an efficient search algorithm for a high speed and high accuracy LVCSR system. A conventionally used lexical tree search is an efficient method, but has a problem in incorporating the language probability. To solve this problem, we propose in this paper a new efficient search algorithm incorporating the language model structure. In our developed LVCSR, 2-pass search algorithm is adopted to produce a word graph as an intermediate expression. The experimental results on the 20,000-word Japanese dictation task showed that the proposed method can reduce approximately half of the processing time without increasing any errors.