采用词图相交融合的语音关键词检测方法

针对词图合并方法产生的词图冗余信息过多，规模较大，导致检索速度较慢的问题，本文提出了一种基于词图相交融合的语音关键词检测方法。首先，将不同语音识别系统产生的词图取交集，并对相同路径上的声学模型、语言模型得分进行得分融合；然后，对于融合后词图中存在的间断路径，直接利用性能最优的语音识别系统产生的词图进行补充，得到完整的融合词图；最后，在相交融合后的词图上进行关键词检测。实验表明，相交融合后的词图综合利用了各词图的得分信息，在基本不损失词图对正确内容覆盖率基础上，减少了冗余信息，有效降低了索引规模；并且在关键词检测性能ATWV指标下，基于词图相交融合的关键词检测方法相比词图合并方法相对提升5.3%。 

Keyword Spotting Using the Lattice Intersection Fusion

In modern keyword spotting implementation, lattice combining is a useful method. Regrettably, it can result in massive redundant information which leads to a slow indexing. This paper propose a keyword spotting method based on lattice intersecting. Firstly, we take the intersection of two lattices, and combining the acoustic score and language score on the same path. Secondly, in order to solve the problem of gaps in the intersection lattice, we use the best single lattice to supplement it. Finally, the intersection lattice will be used for retrieval. With intersection lattice, it utilized the scores of each sub lattice. Without degrading the lattice coverage accuracy, it substantially eliminated redundant information and decreased indexing volume. The experimental results show that keyword spotting can improve 5.3% quality compared to lattice combining referred to keyword spotting ATWV indicator.