消除溢出问题的精确Baum-Welch算法  

Accurate Baum Welch Algorithm Free from Overflow

作　　者：贾宾[1] 朱小燕[2] 罗予频[1] tsinghua.edu.cn 胡东成[1] 

机构地区：[1]清华大学自动化系,北京100084 [2]智能技术与系统国家重点实验室

出　　处：《软件学报》2000年第5期707-710,共4页Journal of Software

基　　金：国家自然科学基金!(No.69982005)资助

摘　　要：Baum- Welch算法是在语音领域中用于 HMM( hidden Markov model)模型参数训练的最基本方法之一 .但它在多样本训练时存在着严重的上、下溢问题 ,需要不断地人工介入来调整中间参数 .该文提出了一种新的能消除上、下溢问题的 Baum- Welch改进算法 .该算法不但摆脱了人工介入 ,保证了计算的精度 ,而且不会带来过大的计算和存储要求 .Baum Welch algorithm, which is often troubled with overflow, is one of the basic methods in the field of speech signal processing. People have to adjust the inner parameters constantly. So in this paper, a modified Baum Welch algorithm is presented to avoid the overflow completely. With this algorithm manual adjustment is not needed and the computation accuracy is guaranteed. There is no significant extra cost of computation and storage. The feasibility of the new algorithm is shown in the experiment.

关 键 词：隐马尔可夫模型 Baum-Welch算法 溢出 语音识别 

分 类 号：TN912.34[电子电信—通信与信息系统]