基于加权有限状态转换器的语音查询项检索技术  

Spoken Term Detection Techniques Based on Weighted Finite-State Transducer

作　　者：陆梨花[1,2] 张连海[1] 陈琦[1] 

机构地区：[1]解放军信息工程大学信息系统工程学院 [2]中国人民解放军95826部队

出　　处：《数据采集与处理》2015年第2期390-398,共9页Journal of Data Acquisition and Processing

基　　金：国家自然科学基金(61175017)资助项目

摘　　要：为了提高语音查询项检索效率,提出了一种在加权有限状态转换器(Weighted finite-state transducer,WFST)框架下以混淆网络代替词格建立索引的技术。在索引建立阶段,首先将词格转化为混淆网络并用自动机形式表示,然后利用自动机构建基于时间的因子转换器,最后将所有因子转换器进行联合及优化得到索引。在查询阶段,将查询项转化为自动机形式后与索引进行合成运算得到表示查询结果的自动机。实验结果表明,在保证系统检测正确率的前提下,与直接以词格建立的WFST索引相比,以混淆网络建立的WFST索引尺寸更小,检索速度更快,因而系统性能更好。An indexing method based on confusion network instead of Lattice is proposed in the weighted finite-state transducer framework（WFST）to improve the efficiency of the spoken term detection system.In the indexing stage,firstly confusion networks are extracted from Lattices and transformed to automatons;Then,timed factor transducers are constructed with these automatons;Finally,the index is achieved by taking the union of the factor transducers and optimizing the union.In the searching stage,the queries are transformed to automatons and then composed with the index.After optimization,the automaton representing the searching results is obtained.Experimental results show that compared with the WFST index based on Lattice,the confusion network-based index has smaller index size,faster searching speed and better performance when ensuring the retrieval accuracy.

关 键 词：加权有限状态转换器 语音查询项检索 混淆网络 因子转换器 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]