语音识别中结合进化计算与MDI的HMM训练方法  

A HMM Training Method Based on Evolutionary Computation and MDI in Speech Recognition

作　　者：何志国[1] 何钦铭[1] 陈奇[1] 曹玉东[2] 

机构地区：[1]浙江大学计算机科学与技术系,浙江杭州310027 [2]攀枝花学院计算机系,四川攀枝花617000

出　　处：《系统工程理论与实践》2006年第7期54-58,65,共6页Systems Engineering-Theory & Practice

摘　　要：提出了将最小差别信息(MDI)和进化计算(EC)相结合引入到HMM的训练中去的方法.各个模型用个体来表示,个体的适应值采用模型的最小差别信息.这样借助于进化计算全局搜索的特点,能克服传统的MDI局部搜索的不足,从而得到系统的全局最优解.实验结果表明,该方法训练所得的系统识别率高于传统的MDI方法训练所得的系统.This paper proposes a training HMM method which based on Minimum Discrimination Information and Evolutionary Computation. Each individual in evolutionary computation roprosents a HMM, while the fitness value of each individual represents the Minimum Discrimination Information. So by means of the global search of EC, this method can overcome the shortcoming of MDI. That is to say, MDI is based on local search. By this method we can gain a global optimum solution. The experimental results indicate that the system＇s recognition rate trained with the proposed method is superior to the one trained with the traditional MDI training method.

关 键 词：最小差别信息 进化计算 隐含马尔柯夫模型 语音识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]