论文部分内容阅读
为了更好的将区分式分类方法应用于说话者确认系统中,该文提出了一种应用于支持向量机(supportvector machine,SVM)说话者确认系统的新型序列核,通过Gauss混合模型训练出每个说话人模型超向量作为支持向量机的输入样本,然后根据Gauss混合模型之间的Kullback-Leibler距离度量构造的SVM序列核函数对超向量进行训练和判决。在美国国家标准与技术研究所(NIST)2004和2006年说话人识别数据库上的实验证明了该核函数能在一定程度上提升整个说话者确认系统的识别精度和鲁棒性。结果表明,本文提出的应用于说话者确认系统中的核函数不仅具有明确的物理意义,而且改善了识别系统的性能。
In order to apply the discriminative classification method to the speaker verification system better, this paper proposes a novel kernel used in speaker verification system of support vector machine (SVM). Gaussian mixture model is used to train every SVM sequence kernel vector based on the Kullback-Leibler distance measure between Gauss mixture models is used to train and decide the supervector. Experiments on the National Institute of Standards and Technology (NIST) speaker recognition database in 2004 and 2006 demonstrate that the kernel function can improve the recognition accuracy and robustness of the whole speaker recognition system to a certain extent. The results show that the proposed kernel function proposed in this paper not only has a clear physical meaning but also improves the performance of the recognition system.