一种基于模型距离的改进的说话人识别技术

The improved speaker recognition system based on model distance measure

提出了一种测量高斯混合模型距离的方法。基于此方法提出一种改进的说话人识别系统:首先从语音中提取几种参数;再分别训练高斯混合模型;然后选择使说话人辨认系统模型平均距离较大的那种特征参数的高斯混合模型,作为该说话人的训练模型;最后在识别时提取此种特征参数进行识别。本文仿真了两种不同模型平均距离的特征参数的正确识别率。实验结果表明:对说话人辨认系统来说,采用使模型平均距离较大的特征参数,所对应的识别性能较好。

This paper presents a distance measure between two Gaussian mixture model. Based on it,an improved speaker recognition system has presented. Firstly, feature is extracted from speech signals. Secondly, Gaussian mixture model of feature is trained. thirdly feature that has the biggest model average distance, is used to recognize. Experiments show that the bigger the model average distance of feature is,the better the performance of system is.