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.
出处
《清华大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
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)
关键词
语音识别
词图
解码
正向-反向算法
speech recognition
lattice
decoding
forward-backward algorithm
