一种快速的语音识别词图生成算法  

Fast lattice generation algorithm

作　　者：李伟[1] 吴及[1] 王智国[1] 

机构地区：[1]清华大学电子工程系,北京100084

出　　处：《清华大学学报（自然科学版）》2009年第S1期1254-1257,共4页Journal of Tsinghua University(Science and Technology)

基　　金：国家"十一五"科技攻关项目;国家"八六三"高技术项目(2006AA01Z149)

摘　　要：词图的高效生成算法是语音识别领域的重要研究课题。该文提出了一种基于词格的词图生成算法(trellis-based lattice-generating algorithm,TBLG),该算法在正向Viterbi解码生成的词格(trellis)基础上,进行反向A*解码生成词图。实验结果表明,与经典的解码器HDecode相比,TBLG生成的词图最优备选效果优于Hdecode。生成高密度词图时,TBLG在解码速度上远远快于HDecode。同时在相同识别率下,TBLG算法生成的词图更加简洁。High efficiency generating lattices is an important area in speech recognition.This paper presentes a trellis based lattice generation algorithm(TBLG).This method uses a backward A* algorithm based on trellises generated by the forward Viterbi decoding to generate the lattices.Experiments show that the lattices generated by this algorithm have better 1-best results than the classical decoder HDecode,the algorithm is much faster than HDecode When generating high-density lattices,and the lattices are much smaller than that of HDecode to get similar performance.

关 键 词：语音识别 词图 解码 正向-反向算法 

分 类 号：TN912.34[电子电信—通信与信息系统]