基于置信度的非特定人语音识别拒识算法的研究  

Study on an Algorithm of Speaker-independent Utterance Rejection Based on Confidence Measures

作　　者：贲俊[1] 万旺根[2] 余小清[1] 

机构地区：[1]上海大学通信与信息工程学院,上海200072 [2]香港理工大学多媒体创新中心

出　　处：《计算机应用研究》2003年第7期16-19,共4页Application Research of Computers

基　　金：国家自然科学基金资助项目(60072031);香港理工大学研究基金资助项目(G T499)

摘　　要：语音拒识技术是实现一个实用语音识别系统的关键。提出了一种新颖的基于置信度的非特定人语音识别拒识算法,该算法同时考虑了备选假设模型和多候选的信息,适用于拒识不正确的识别结果和词表外(OOV)语音。在一个非特定人英语命令词识别系统中做了一些相关的实验来评估这个算法的性能。实验结果表明,该算法可以有效地去除识别不可靠的语音,提高语音识别的整体性能。Utterance rejection is the key technology in the realization of practical speech recognition system. In this paper a new algorithm of speakerindependent utterance rejection based on confidence measures is proposed, which takes alternative hypothesis model and multicandidates into account and is suitable for rejecting wrong recognition result and Out Of Vocabulary(OOV). Experiment is conducted in a speakerindependent English command recognition system to evaluate the algorithm and the results show that the algorithm can reject incorrect recognition results and improve the whole performance of speech recognition.

关 键 词：置信度 语音识别拒识 隐马尔可夫模型 

分 类 号：TN912.32[电子电信—通信与信息系统]