语音识别中空间相关性信息的利用

在语音识别中不同语音单元的语音信号特征之间不是独立的,描述不同声学状态信号特征之间的相互关系的信息称为"空间相关性".空间相关信息在语音识别参数估计算法(训练和自适应)中有非常重要的作用.本文对这种相关性作了探讨,提出了一种在语音识别中应用空间相关信息的新方法.我们用线性方程来描述空间相关性所体现出来的不同语音单元特征之间的依赖性,通过分组K-L变换的方法来估计这组线性约束的相关系数,并给出一种结合空间相关信息的训练方法.实验结果表明,空间相关的先验知识对语音识别训练模块的稳健性有明显的提高.

Using Spatial Correlation Information in Speech Recognition

In Speech Recognition,features from different acoustic units are not independent.The correlation between different acoustic units is called "Spatial Correlation".Spatial Correlation is important for acoustic model estimation.In the paper,we proposed a new method of using spatial information in speech recognition.We use linear equation to subscribe spatial correlation,calculate equation coefficients by K-L transformation,and develop a new training algorithm with the linear constraints.Experimental results show the new method brings significant improvement in error reduction.