为了提高语音识别系统的性能，基于全域优化的思想，提出了一种用于训练连续隐马尔柯夫模型（ＣＨＭＭ）的新算法——基因算法，并将该算法用于语音识别．用该算法训练ＣＨＭＭ，可得到最佳的模型参数，从而提高了语音识别率．利用该算法训练ＣＨＭＭ，不需要对ＣＨＭＭ的每一个参数单独进行估值，能够在一定的程度上提高训练速度．文中阐述了整个算法，给出了计算机模拟结果，并与传统的训练方法进行了比较． In order to improve the performance of speech recognition system, a new algorithm for training continuous hidden Markov model (CHMM) -gene algorithm is proposed based on the idea of ​​global optimization. The algorithm is applied to speech recognition. Using this algorithm to train the CHMM, the best model parameters are obtained, which improves speech recognition rate. Using this algorithm to train CHMM, it is not necessary to evaluate each parameter of CHMM separately, which can improve the training speed to a certain extent. In this paper, the whole algorithm is expounded, the computer simulation results are given, and compared with the traditional training methods.