基于KLD差的统计错误模式生成算法  

KLD Difference-Based Statistical Error Pattern Generation for Pronunciation Quality Assessment

作　　者：刘庆升[1] 魏思[1] 胡郁[1] 王仁华[1] 

机构地区：[1]中国科学技术大学讯飞语音实验室,合肥230027

出　　处：《数据采集与处理》2009年第1期32-37,共6页Journal of Data Acquisition and Processing

基　　金：国家语言文字工作委员会"十五"重点科研(ZDI105-B02)资助项目

摘　　要：研究了用于指导计算机发音质量评价的错误模式的生成算法,它是普通话CALL系统研究工作中的一部分。传统的错误模式是根据语言学知识来生成的,只能得到那些最重要的常见错误模式。为了提高错误模式的覆盖面,本文提出了一种基于KLD差的统计错误模式生成算法,用模型间KLD作为模型间的距离,以标准模型间KLD与带方言口音模型间KLD的差代表两种模型间的差异,并以之为度量来生成错误模式。实验证明在引入了此算法生成的错误模式后,系统性能由0.809提升到0.826。The generation algorithm for the error pattern is studied to guide the pronunciation quality assessment. The error pattern refers to rules that one phone was pronounced into another wrong one due to the influence of the speakers accent. The error pattern can be applied in CALL system to generate a compact recognition network in which the recognizer errors are reduced and the system performance is improved. Error patterns are used from linguistic knowledge, which covers the most familiar error patterns. This paper presents a KLD（Kullback-Leibler divergence） difference-based error pattern generation algorithm and the algorithm uses KLD as the distance between two models. This new algorithm measures the differences between KLDs calculated from standard models and KLDs from the accented models to generate the error pattern. Experiments prove that with the error pattern generated by the new algorithm, the system performance is improved from 0.809 to 0.826.

关 键 词：语音识别 中文信息处理 发音质量评价 KLD 

分 类 号：TN912.34[电子电信—通信与信息系统]