Two hybrid training methods, GA forward-based GA-HMM model and GA embed-based GA-HMM model, are proposed. GA can be used to adjust the initial training parameters of HMM randomly. It can avoid falling of the training of HMM into local optima effectively.

• 为此,采用GA训练HMM模型参数,并给出了GA和HMM的两种混合训练方式:前端GA HMM模型和内嵌式GA HMM模型,GA算法能随机地调整HMM模型训练的初始值,使HMM跳出局部最优,较好地克服了HMM训练容易陷入局部最优的问题。