基于环境特征判别学习的顽健语音识别方法

摘要

A discriminative learning method of environmental features isproposed for robust speech recognition,in which a simple classifier and the gradient descent algorithm are firstly used to iteratively learn the environmental features,and then the pure speech features are estimated from the observed speech features.Finally the estimations of the pure speech features are used in the back-end HMM classifier.Using the proposed method,a 33.3% reduction in word error rate is obtained relative to conventional HMM system for a speaker-independent,small vocabulary recognition task.%提出一种基于环境特征判别学习的顽健语音识别方法,它首先通过使用一个简单的分类器和梯度下降法迭代地学得环境特征,接着利用得到的环境特征从观测到的混噪语音特征中估计出纯净的语音特征,然后将估计出来的纯净语音特征用到后端的HMM分类器中.使用所提出的方法对不特定话者小词表进行实验,其系统误识率与基本HMM系统相比下降了33.3%.