基于得分归一化和系统融合的语音关键词检测方法  

Keyword Spotting based on Score Normalization and System Combination

作　　者：李鹏[1] 屈丹[1] Li Peng Qu Dan(School of Information Systems Engineering, PLA Information Engineering University, Zhengzhou, 450001, China)

机构地区：[1]解放军信息工程大学信息系统工程学院,郑州450001

出　　处：《数据采集与处理》2017年第2期346-353,共8页Journal of Data Acquisition and Processing

基　　金：国家自然科学基金(61175017)资助项目

摘　　要：为了有效利用不同关键词检测系统的互补性,解决不同系统检测结果置信度得分不在同一范围的问题,提出了一种基于得分规整和系统融合的语音关键词检测方法。首先,为了克服连续语音识别系统中因剪枝错误而引起的关键词丢失问题,应用了关键词相关的软Beam宽度剪枝策略裁剪词图;其次,在系统融合前采用得分归一化方法,使得不同系统关键词检测结果置信度得分在同一范围;最后,通过系统融合处理将不同系统的关键词输出进行整合,得到最终的关键词检测结果。实验结果表明,经过得分归一化处理后,关键词检测性能的实际查询词权重代价(Actual term-weighted value,ATWV)平均相对提升30%;系统融合后关键词的检测性能,相比于得分归一化处理后的最佳单一系统,得到了10%的提升。To effectively use the complementarity of different keyword spotting systems and solve the problem that the confidence scores from several different subsystems is not in the same range, a keyword spotting system based on score normalization and system combination is proposed. Firstly, to avoid keyword missing due to pruning errors in a large vocabulary recognition system, the keyword soft Beam pruning method is presented. Secondly, score normalization is needed to transform these confidence scores into a common domain, prior to combining them. Finally, after score normalization,the outputs are combined from different subsystems. Results show that score normalization methodology improves keyword search performance by 30 % in average. Experiment also show that combining the outputs of di- verse systems, system perform is 10% better than the best normalized KWS system.

关 键 词：关键词检测 得分归一化 系统融合 软Beam剪枝 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]