说话人辨认中基于进化策略的最大互信息训练方法

针对最大似然训练分辨能力的不足，把最大互信息训练方法引入到高斯混合模型（GMM）的训练中，并直接采用进化策略实现模型参数的全局训练，以模型与训练数据之间的互信息作为进化过程中个体的适应度。该系统不仅分辨能力强，而且摆脱了局部搜索的缺陷。实验结果表明，这种方法生成的说话人辨认系统的识别性能要优于传统的期望最大化算法（EM）生成的系统。

Maximum Mutual Information Training Based on Evolution Strategy in Speaker Identification

Due to the innate weakness of maximum likelihood training, a maximum mutual information based training was introduced into the training of Gauss mixture models(GMMs) and the evolution strategy is applied directly to implement the training. During the evolution the mutual information between the models and the training data is used to evaluate the fitness of an individual. The experimental results indicate that the speaker identification system is superior to the traditional expectution maximization (EM) algorithm.