最小错识率（MCE）HMM训练方法是最直接的判决训练方法之一，原理上比最大似然接方法优越得多。但是，充分发挥MCE的性能的困难在于如何将非连续的经验错识率估计映射成连续的代价函数以便采用梯度搜索技术之类的忧化方法。本文所提出的基于遗传算法的最小错识率方法（GAMCE）不但可避免MCE的缺点，而且在原理上可求得全局最优解。实验结果表明，GA_MCE比MCE好得多。 Minimum error rate (MCE) HMM training method is one of the most direct decision training methods, in principle, is much superior to the maximum likelihood method. However, the difficulty of giving full play to the performance of MCEs lies in how to map discontinuous empirical error rate estimation as a continuous cost function in order to adopt a mitigation approach such as gradient search techniques. The GAMCE based on genetic algorithm proposed in this paper can not only avoid the shortcomings of MCE, but also get the global optimal solution in principle. The experimental results show that GA_MCE is much better than MCE.