利用不匹配帧比例加权N-Best候选的置信度方法

针对关键词检出中的置信度可靠性问题,提出了一种基于N-Best候选路径的置信度方法,根据比较候选路径上关键词边界内不匹配帧所占的比例大小,对N-Best的结果进行加权.该方法不但利用了词图搜索上N-Best候选的似然分,而且突出了每条路径中局部匹配的置信度,在上下文相关的中文关键词检出系统上,通过重庆口音普通话数据库的...

Approach of N-Best lists-based confidence measure-using mismatching frame proportion weighting

In resolving the reliability of confidence measure for Keyword spotting, an N-Best lists-based Confidence Measure is proposed which weights and combines each N-Best list's likelihood score by comparing the proportion of Mismatch Frames dominated by hypothetical segment and by the whole sentence. This approach not only utilized the likelihood scores of N-best alternatives, but also emphasized the confidence within local segment in each list. Over the Chongqing accented dialectal database, the Equal Error Rate (EER) of a context dependent phonetic keyword spotting system implemented with this confidence measure has been reduced by 14.2% relatively compared with the traditional Duration Normalized Likelihood baseline.