首先讨论了基于MCE/GPD的语音识别研究的最新进展。在此基础上 ,提出了一种环境特征判别学习的Robust语音识别方法 ,该方法基于最小分类错误准则利用梯度下降法迭代地学习环境特征。由于梯度下降法产生的是局部最优解 ,因此 ,寻找较好的环境特征初始值就显得非常重要。最后 ,讨论了这种环境特征判别学习方法中参数的初始值选择问题 First of all, the latest research on speech recognition based on MCE / GPD is discussed. On this basis, Robust speech recognition method based on discriminative learning of environmental features is proposed, which uses the gradient descent method to iteratively learn the features of the environment based on the minimum classification error criterion. As the gradient descent method produces a local optimal solution, it is very important to find the initial value of a good environmental feature. Finally, the initial value selection of parameters in this method of environmental feature discrimination is discussed