基于多核SVM-GMM的短语音说话人识别

Speaker recognition with short utterances based on multiple kernel SVM-GMM

运用多个核函数的线性组合构造多核空间,在多核空间上设计了基于支持向量机的说话人分类器,实现短语音说话人识别。多核映射能够解决单核映射核函数及其参数选择的难题,增加说话人的可区分性,提高分类器的性能。算法中结合了高斯混合模型(GMM),并以GMM超向量作为说话人的最终特征参数进行仿真实验。实验表明,在短语音和两种噪声环境中,基于多核SVM-GMM的短语音说话人识别算法较SVM-GMM算法能得到更好的识别性能和鲁棒性。

A linear combination of several kernels is used to construct multiple kernel space. In multiple kernel space, Support Vector Machine （SVM） classifiers are designed to identify speakers with short utterances. Multiple kernel mapping can solve the problem of single kernel mapping, such as the selection of kernel function and parameters. Besides, multiple kernel mapping can increase discriminative power among different speakers and improve the performance of classifiers. In simulation experiment, Gaussian Mixture Model （GMM） was used to get GMM supervector as speakers＇ final feature parameters. Experiment results show that under the condition of short utterances and two noisy environments, the performance and robustness of the multiple SVM-GMM speaker recognition algorithm are better than that of SVM-GMM algorithm.