语音关键词检测系统中基于时长和边界信息的置信度  

Confidence Measure Based on Time and Boundary Features for Speech Keyword Spotting System

作　　者：李文昕[1] 屈丹[1] 李弼程[1] 王炳锡[1] 

机构地区：[1]解放军信息工程大学信息工程学院,郑州450002

出　　处：《应用科学学报》2012年第6期588-594,共7页Journal of Applied Sciences

基　　金：国家自然科学基金(No.61175017;No.60872142)资助

摘　　要：针对关键词识别中基于后验概率的置信度方法对语音发音变化信息利用不充分的问题,提出了一种基于时长和边界信息的置信度.该方法引入一个松弛因子,灵活地选择词信息相同的弧段计算置信度,从而进行关键词拒识.在此基础上,设计并实现了一个基于Lattice的大词表语音关键词检测系统,先用改进的动态规划算法在音节网格上进行关键词检出,尽可能多地给出关键词候选,然后采用基于时长和边界信息的置信度进行关键词确认.实验结果表明,该方法相对于主流的计算方法,系统的等错误率(equal error rate,EER)提高了7%.As the posterior probability confidence measure cannot take full advantage of the variations in voice pronunciation, we propose an improved confidence measure based on time and boundary feature. A relaxation rate is introduced to have a flexible selection of the segmental arcs with the same words for the calculation of confidence by which detections are rejected. On this basis, a keyword spotting system with large keyword table based on lattice is designed. An improved dynamic time warping （DTW） algorithm is used for keyword matching through the lattice to generate maximum keyword hypotheses. We have applied the proposed confidence measure in keyword verification. The results show that the equal error rate （EER） achieves 7% relative improvement compared to the mainstream method of calculation.

关 键 词：语音识别 关键检出 置信度计算 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]