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 ,避免了溢出问题 In the training phase of HMM model of speech recognition system, due to the large number of continuous products in the Baum Welch algorithm, the numerical results will be smaller and smaller, resulting in overflow. In the case of single observation sequence calibration Technology can properly solve the problem of spillover.In the case of multiple observation sequences, the output probability of each sequence to the HMM is introduced as a correction coefficient, the value is small, the problem of overflow still exists.This paper analyzes the reasons for the overflow problem, Observe the situation of the sequence, and change the optimization objective function from the product multiplication of the output probability to the logarithmic summation form, and deduce a set of improved Baum Welch algorithm. The algorithm reduces the computational complexity of the HMM parameter revaluation algorithm, improves the stability and avoids the overflow problem