一种改进动态特征参数的话者语音识别系统

A Speaker Voice Recognition System of Improved Dynamic Characteristic Parameters

研究语音动态特征参数提取问题,在话者语音识别过程中,动态特征参数可以有效提高识别率。但是传统算法在其提取过程中存在大量干扰冗余信息,造成了识别率降低并带来运算速度的降低。为解决上述副作用,提出在说话人识别系统中,使用一种动态时频倒谱系数参数的方法。上述方法在不减少反应话者个体特征分布特性的前提下,可消除冗余信息并降低样本特征的维度。利用上述方法提取语音特征参数并输入混合高斯-通用背景模型进行说话人语音分类。在Matlab上仿真结果表明,动态时频倒谱系数可有效改进话者语音识别系统的识别正确率。

In Speech Recognition System（ SRS）, the dynamic feature parameters can effectively improve the recognition rate. Traditional algorithm can generate redundant information in the extraction process, which reduce the recognition rate and arithmetic speed. In order to solve the above side effects, this paper presents a parameter extraction method based on dynamic timing frequency cepstral feature （DTFC） in the SRS. The improved method can eliminate the redundant information and reduce the feature dimension in the premise of not to reduce distribution characteristics of the individual as far as possible. In addition, the SRS uses the Gaussian Mixture Model - Universal Background Model（ GMM - UBM） for speaker speech classification. Matlab simulation results show that the DTFC can increase the speech recognition accuracy.