中文摘要：

为了得到更具区分性的特征参数,采用改进的MFCC提取方法,即低方差性的多窗谱估计MFCC,并在其基础上引入了短时TEO能量和ΔMFCC动态特征参量组合特征进行说话人识别。由于直接将两者进行组合会造成维度过高,计算复杂度增加,为此提出了相关距离Fisher比来对特征参数进行加权和维度筛选,最后送入GMM-UBM分类器模型进行识别。实验表明,改进的混合特征参数相较于单一的特征参量,具备更好的识别能力,使得识别率有一定程度的提高。

英文摘要：

In order to get more distinguished characteristic parameters, we utilize a improved multitaper MFCC extraction algorithm which with low variance. On the basis of this, we propose mixed characteristic parameters which combined short-time TEO energy with first-order MFCC that time-domain characteristics and reflecting individual voice dynamic characteristics of the speech signal separately. Due to combing the two parameters directly will result in high dimension parameters and increase the complexity of computational, for this reason, we propose a algorithm for feature selection about fisher criterion with correlation distance. Then, the speaker recognition is based on GMM-UBM classification model. ExperimentS show that the improved mixed characteristic parameter compared to single characteristic parameters has better recognition results and improving the system recognition rate.