用于连续语音识别的词汇搜索算法

给出一种用于连续语音识别的词汇树搜索算法．这种方法基于启发搜索，分为音节层—树搜索、词层—树搜索两个过程．为在提高识别速度的同时不牺牲识别率，上述两个过程采用的是“有限宽度树搜索”．在音节层—树搜索中，根据Ｎｂｅｓｔ算法确定允许扩展的有限个音节串．而在词层—树搜索中，根据已得到的允许扩展的音节串确定不同长度的“退回词路径”．这样，每一时刻的最大似然词串可由这些“退回词路径”来确定．这种算法的优点是在搜索不同长度的多音节词的同时，可以减少搜索计算开销和降低对存储空间的需求．

A Lexical Access Strategy for Continuous Speech Recognition

Presents a lexical tree search strategy for continuous speech recognition. This algorithm is a variant of heuristic search strategy, with two level tree beam search processes which divided into syllable level tree search process and a word level tree search process. The beam search can reduce the necessary of computation without sacrificing the accuracy.In the syllable level search process the admissible syllable sequences are explored using the N best algorithm.And in the word level search process the backtracking word paths are explored from those admissible syllable sequences. Hence, the best current backward word path ending at each node in the word level tree can be determined with these backtracking word paths. This given strategy has the merit of exploring multi syllable words with the reduction of computation and memory necessary.