基于加权有限状态机的动态匹配词图生成算法

发布时间：2018-09-05 17:14

【摘要】：由于现有的加权有限状态机(WFST)解码网络没有精确词尾标记,导致当前已有的词图生成算法不含精确的词尾时间点,或者仅是状态、音素级别的词图,无法应用到关键词检索中。该文提出在WFST静态解码器下的语音识别词图生成算法。首先从理论上分析了WFST解码音素图和词图的可转换关系,然后提出了字典的动态音素匹配方法解决了WFST网络中词尾时间点对齐的问题,最后通过令牌传递的遍历方法生成了词图。同时,考虑到计算量优化,在令牌传递过程中引入了剪枝算法,使音素图转词图的耗时不到解码耗时的3%。得到的词图,不仅可以用于语言模型重打分,由于含有精确的词尾时间点,还可以直接应用到关键词检索系统中。实验结果表明,该文的词图生成算法具有较高的计算效率;和已有动态解码器的词图相比,词图中包含更多解码信息,在大词汇连续语音识别的重打分结果和关键词检索中都能取得更好的性能。
[Abstract]:Because the existing weighted finite state machine (WFST) decoding networks do not have accurate endings, the existing word graph generation algorithms do not contain accurate word end time points, or only state, phoneme level word images, which can not be applied to keyword retrieval. A speech recognition word graph generation algorithm based on WFST static decoder is proposed in this paper. In this paper, the convertible relation between WFST decoded phoneme graph and word graph is analyzed theoretically, and then the dynamic phoneme matching method of dictionary is proposed to solve the problem of word end point alignment in WFST network. Finally, the word graph is generated by the traversal method of token passing. At the same time, considering the computational optimization, a pruning algorithm is introduced in the token passing process, which makes the conversion time of phoneme graph less than 3 times of decoding time. The obtained word graph can not only be used for rescoring the language model, but also can be directly applied to keyword retrieval system because of the precise time point of the end of the word. The experimental results show that the algorithm has a high computational efficiency and contains more decoding information than the word graph of the existing dynamic decoders. Better performance can be obtained in rescoring and keyword retrieval of large vocabulary continuous speech recognition.
【作者单位】：中国科学院语言声学与内容理解重点实验室;
【基金】：国家自然科学基金(10925419,90920302,61072124,11074275,11161140319,91120001,61271426) 中国科学院战略性先导科技专项(XDA06030100,XDA06030500) 国家863计划项目(2012AA012503) 中科院重点部署项目(KGZD-EW-103-2)资助课题
【分类号】：TN912.34;TP301.1

【参考文献】