Abstract: The classical Baum Welch algorithm for Hidden Markov Model（HMM） training is easily trapped in local optimum.To this question，this paper proposes a gene cloning Baum Welch algorithm.When the result of Baum Welch algorithm changes less than 10－3，based on current local optimized parameters matrix B，it executes the gene cloning operator to get optimized HMM parameters matrix B.At last，the probability of voice templates of Baum Welch algorithm output is improved.Experimental results show that the template probability of new algorithm is greater than the classic Baum Welch algorithm，and the new training template is better than Baum Welch algorithm’s.

Key words: speech training, Hidden Markov Model（HMM）, Baum Welch algorithm, gene cloning

摘要： 针对用于隐马尔科夫模型（HMM）训练的经典Baum Welch算法容易陷入局部最优解这一问题，提出基因克隆的Baum Welch算法。该算法在Baum Welch算法迭代计算到10－3以内不再改变的情况下，在当前已获得局部最优参数B矩阵的基础上，执行基因克隆算子，获得优化的HMM的B参数，进一步提升Baum Welch算法语音模板的输出概率。实验结果表明：该算法模板计算概率大于经典的Baum Welch算法，获得了比Baum Welch算法更优的训练模板。

关键词: 语音训练, 隐马尔科夫模型, Baum Welch算法, 基因克隆