One-pass LVCSR algorithm using linear lexicon search and 1-best approximation tree-structured lexicon search

We propose a combinational use of linear lexicon and tree-structured lexicon in a 1-best approximation search algorithm for large vocabulary continuous speech recognition. The algorithm based on 1-best approximation with a tree-structured lexicon is efficient but frequently looses the optimal sentence hypothesis. The linear lexicon search can find the optimal hypothesis but needs much computational cost. Thus, we propose a search method using these two search algorithms in parallel to achieve efficient and accurate decoding. We also adopt trigram and 4-gram language models to this algorithm to achieve 1-pass tri-gram and 4-gram decoding. These combination looses optimal hypotheses, but the usage of the linear lexicon reduced such errors. We introduce dasialikelihood difference indexpsila to realize accurate inter-word context-dependent modeling. We evaluated this new search algorithm and obtained significant improvement of recognition performance without severe increase of computational cost.