语种识别算法中GSV计算的定点仿真与实现  

Fixed-point simulation and realization of GSV used in language recognition

作　　者：张丽[1] 杨镇西[1] 吉立新[1] 

机构地区：[1]国家数字交换系统工程技术研究中心,河南郑州450002

出　　处：《计算机工程与设计》2012年第2期679-683,共5页Computer Engineering and Design

基　　金：国家863高技术研究发展计划基金项目(2008AA011002)

摘　　要：基于GSV-SVM的语种识别方法是目前最为流行的语种识别方法之一,其采用基于通用背景模型GMM-UBM的GSV作为声学模型,支持向量机SVM作为区分模型。大量仿真测试结果表明,GSV在整个系统中占的运算量为80%左右,是算法硬件实现的瓶颈。鉴于此,对基于GSV的硬件实现方法进行了研究,提出了一种快速GSV定点计算方法,其采用addlog运算简化对数似然函数的计算,完成了语种识别的高效定点实现。实验结果表明,该定点方法的识别率与浮点识别基本一致,满足应用要求。The language recognition method based on GSV-SVM is the most popular language recognition method.In this method,the GSV（GMM Supervector） based on GMM-UBM（Gaussian mixture model-universal background model） is used as the acoustic model,and SVM（support vector machine） is used to distinguish the model.After a lot of simulations,the computing capacity of GSV account for 80 percent in the system,it＇s the bottleneck of the algorithm for hardware implementation.Based on that,Hardware implementation method of GSV is studied,fixed-point method is proposed,and the addlog operation is adopted to simplify the log-likelihood function calculation,and an efficient fixed-point implementation of language identification is achieved.The result of experimentation shows that the recognition ratio of fixed-point method is consistent with that of floating point method,and meets the application requirements.

关 键 词：语种识别 高斯混合模型-通用背景模型 GMM超矢量 定点实现 addlog运算 

分 类 号：TP391.42[自动化与计算机技术—计算机应用技术]