基于互信息置信度的网格连续汉语语音检索  

Lattice-based continuous Chinese speech indexing based on mutual information confidence measure

作　　者：黄湘松[1] 赵春晖[1] 张磊[1] 刘柏森[2] 

机构地区：[1]哈尔滨工程大学信息与通信工程学院,哈尔滨150001 [2]黑龙江工程学院电子工程系,哈尔滨150050

出　　处：《计算机应用研究》2009年第12期4607-4609,4616,共4页Application Research of Computers

基　　金：国家自然科学基金资助项目(60702053)

摘　　要：针对目前生活中涌现的海量语音数据,人们对语音检索技术准确度的要求越来越高。主要研究了汉语连续语音检索任务中,基于转换音节网格的研究方法。针对语音检索系统中置信度计算的问题,提出了一种基于音节间互信息的置信度计算方法,并将其用于网格结构的语音检索系统中。该方法能够有效地利用上下文之间的互信息量,从而更准确、合理地描述汉语语言模型。实验结果表明,用提出的方法建立转换音节网格来进行语音检索,其检出率(FOM)比后验概率法和N-best法有较大幅度的提高。得到的汉语语音检索系统其FOM最高可以达到83.7%。Nowadays, with the overwhelming amounts of speech data rushing in our life, higher and higher accuracy of speech indexing techniques is required. This paper mainly studied a converted syllable lattice-based approach in a Chinese continuous speech indexing task. Aiming at the computation of confidence measure in a speech indexing system, this paper proposed a confidence measure method based on mutual information between syllables, which was used in a lattice construction system for speech indexing. The method took full advantage of the context mutual information, which could describe Chinese language model more exactly and logically. The experiment results show that using the proposed method to build a converted syllable lattice in a speech indexing system, the FOM of which has great improvement comparing with posterior probability based method and N-best based method. This best system for Chinese speech indexing achieves a FOM of 83.7%.

关 键 词：网格 互信息 语音检索 置信度 语言模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]