基于新序列核支持向量机的说话人识别

Speaker Recognition Based on New Sequence Kernel SVM

为了更好地将区分式分类方法应用于说话者确认系统中,构建序列核支持向量机已成为说话人识别领域的研究热点与趋势。本文在研究可再生希尔伯特空间框架的基础之上构建出一个新的序列核来对语音序列间的相似性进行度量,并结合近年来提出针对支持向量机(SVM)跨信道子空间特征差异(ISV)所提出的归整技术(LFA,NAP,CSP),进一步优化序列核系统。在美国国家标准与技术研究所(NIST)2004年评测数据集的实验中,新序列核系统的识别率高于传统高斯混合模型(GMM)和基于广义线性区分性核(GLDS)的支持向量机。

To apply the discriminative classifier in the speaker recognition, the building sequence kernel support vector machine （SVM） becomes the trend in the field. By using the framework of the reproducing kernel Hilbert space, a new sequence kernel for measuring the similarity between observations sequences is developped. By combining the later cross-channel compensation technologies （LFA, NAP, CSP）, aimed at intersession variability （ISV）, the sequence system is further optimized. The test evaluation database by the National Institute of Standards and Technology（NIST） 2004 demonstrates that the performance of the new sequence kernel system is superior to that of traditional Gaussian mixture model （GMM） and GLDS SVM.