摘要： 从假设检验和模式识别的角度分别介绍了置信度的概念及其在实际关键词检测系统中的应用，并提出3种新的基于识别结果本身信息的声学置信度：基于状态长度归一化的置信度、基于特征矢量信息的置信度和基于状态驻留的声学置信度。与传统的基于似然比的声学置信度相比，这3种置信度几乎不需额外的训练和建模，计算简单，且均较易设置统一的拒识门限。实验结果表明基于状态驻留和基于特征矢量信息的置信度的性能较好，能够有效降低基线系统的虚警率。

关键词: 关键词检测, 确认, 置信度, 置信区间

Abstract: This paper mainly discusses the application of acoustic confidence measures in keyword verification systems. First it introduces the concept of confidence measure from the view of hypothesis testing and pattern recognition , then brings forward three new confidence measures based on recognition results: the confidence measure based on normalized state length, the confidence measure based on state duration and the confidence measure based on the information of feature vector. Experimental results show that the three confidence measures can get better verification performance and more efficiency compared with the traditional confidence measure based on log likelihood ratio and the threshold of verification can be set uniformly.

中图分类号: 

• TP391.4