一个快速说话人识别系统的设计和实现

The Application of a Fast Multi-language Recognition System

基于美国国家科学技术标准局的说话人识别评测任务,提出了快速说话人识别技术框架。在此框架下,低层的声学特征向量首先经过高斯混合建模和非线性映射,转变为高层的高维特征向量(超向量);接着利用区分性分类器支持向量机,对超向量进行分类;最后依据假设检验理论,进行最终判决。相比较传统的说话人识别系统,以超向量作为特征进行分类,比直接采用声学特征更为稳健。实现了一个快速说话人识别系统,并在电话数据集上进行测试,取得了不错的效果。

We proposed a fast language recognition framework based on National Institute of Standards and Technology Speaker Recognition Evaluation （NIST SRE）. In the framework, low-level acoustic feature vector is firstly converted to high -level feature vector （Super Vector） through Gaussian Mixture Modeling and nonlinear mapping methods. Then, the super vector is classified by backend classifier. Finally, decision is given out based on hypothesis test theory. The proposed system is much more robust compared with traditional speaker recognition system. Besides, a fast speaker recognition system is built and achieved a good result based on telephone dataset.