多观察序列连续隐含马尔柯夫模型的无溢出参数重估  

Re-estimation of Continuous Hidden Markov Model with Multiple Observation without Overflow

作　　者：何强[1] 毛士艺[1] 张有为[2] 

机构地区：[1]北京航空航天大学电子工程系,北京100083 [2]五邑大学信息科学研究所,广东江门529020

出　　处：《电子学报》2000年第10期98-101,共4页Acta Electronica Sinica

基　　金：广东省自然科学基金!(No.960 631 )

摘　　要：在语音识别系统的HMM模型训练阶段 ,由于Baum Welch算法中前向概率和后向概率包含大量连乘项 ,计算结果数值会越来越小 ,以致产生溢出 .在单观察序列情况下采用定标技术可以妥善地解决溢出问题 .在多观察序列情况下 ,则会引入各序列对HMM的输出概率作为修正系数 ,其数值很小 ,溢出问题仍存在 .本文分析了溢出问题产生的原因 ,针对多观察序列的情况 ,将优化目标函数由输出概率的连乘改为对数累加和形式 ,推导出一套改进的Baum Welch算法。该算法降低了HMM参数重估算法的计算复杂度 ,提高了稳定性 。In the training phase of HMM system,due to the cumulative production,the evaluation of the forward and backward probabilities in Baum Welch algorithm needs a large dynamic range which will exceed the precision range of essentially any machine.This can be resolved by multiplying the forward and backward probability with scaling coefficients in single observation case.In multiple observation case,generally the output probability terms are introduced for each observation,which will cause the overflow problem again.In this paper,the cause of overflow problem is studied and a revised BW algorithm specially for multiple observation is deduced by redefining the optimization object function from cumulative production to sum of logarithms.The revised HMM parameter re estimating algorithm is more stable and effective,and the overflow problem is eliminated.

关 键 词：马尔柯夫过程 Baum-Welch算法 话音识别 HMM模型 

分 类 号：TN912.3[电子电信—通信与信息系统]