融合后验概率置信度的动态匹配词格检索∗

摘要

In the keyword spotting system based on dynamic match lattice spotting ( DMLS) , the minimum edit distance is used as the confidence measure. When the detection rate is increased, the false alarm rate is raised as well. Aiming at this problem, an approach integrating the posterior probability confidence measure with DMLS is proposed. Firstly, the posterior probability based on lattice is introduced with the index stage of DMLS. Secondly, data driven phone substitution, insertion and deletion costs are incorporated for more flexible phone sequence matching. Finally, the minimum edit distance and the posterior probability confidence measure are blended together to detect all occurrences of the keywords. The experimental results show that there is a certain complementarity between the minimum edit distance and posterior probability confidence measure and the equal error rate is relatively reduced.%在基于动态匹配词格检索( DMLS)的关键词检测系统中，应用最小编辑距离作为关键词检出的置信度，在提高检出率的同时也增加虚警率。针对此问题，文中提出融合后验概率置信度的动态匹配词格检索方法。该方法首先将基于Lattice的后验概率引入到DMLS的索引建立中，其次应用数据驱动的音素替换、插入和删除代价，实现更灵活的近似匹配，最后通过联合最小编辑距离和后验概率置信度得分进行关键词检测。实验表明，最小编辑距离和后验概率置信度具有一定的互补性，系统的等错误率相对降低。