一种基于SSM的HMM训练算法

一种基于SSM的HMM训练算法 A HMM Training Algorithm Based on SSM 王新民 WANG Xin-min~1 1 2 3 first-author 姚天任 YAO Tian-ren~2 1 2 3 孝感学院物理系 孝感学院物理系 Department of Physics, Xiaogan University Department of Physics, Xiaogan University 华中科技大学电子信息工程系,湖北,孝感,432100 华中科技大学电子信息工程系,湖北,孝感,432100 Department of Electronic and Information Engineering,Huazhong University of Science and Technology Department of Electronic and Information Engineering,Huazhong University of Science and Technology ,湖北,武汉,430074 ,湖北,武汉,430074 在Baum Welch（BW）算法的基础上提出了一种基于态相关方法（State SpecificMethod:SSM）的隐马尔可夫模型（HiddenMarkovMode:HMM）参数估计算法（简称SBW算法）.该算法在估计HMM不同状态的概率密度函数（proba bilitydensityfunction:PDF）的参数时使用了与状态有关的维数较低的特征集合.与传统的BW算法相比,新算法避免了直接估计高维的PDF参数.仿真实验表明,在训练数据量不足的情况下,采用SBW算法的误识率明显低于BW算法. In this paper, we derive an algorithm based on well-known BW algorithm for estimating the parameters of a hidden Markov model （we call it SBW algorithm）. SBW algorithm relies on a low dimensional state-specific feature set rather than rely on a common high dimensional feature set as conventional BW algorithm, so avoiding directly estimating high-dimensional probability density functions in HMM training. Our computer simulation example shows that the performance of the new algorithm is superior （over） the conventional Baum-Welch algorithm. 隐马尔可夫模型 BaumWelch算法 参数估计 态相关方法 充分统计 hidden Markov model（HMM） Baum-Welch algorithm parameter estimation state-specific method sufficient statistics TP301.6 湖北省教育厅重点项目基金资助项目（2002A02004） 2003-05-01 2021-04-01 5