论文部分内容阅读
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 .
In speech recognition, the characteristics of speech signals of different phonetic units are not independent, and the information describing the interrelationships among the signals of different acoustic states is called “spatial correlation.” Spatial correlation information is used in the speech recognition parameter estimation algorithm Adaptive) .This paper discusses the correlation and proposes a new method of applying spatial related information in speech recognition.We use the linear equation to describe the difference of spatial correlation And the dependence between the features of speech units, the group KL transform method is used to estimate the correlation coefficient of the linear constraints, and a training method combining the spatial correlation information is given.The experimental results show that the spatial correlation of prior knowledge of speech The robustness of the recognition training module has been significantly improved.