In text-independent speaker verification, how to extract typical feature which is suitable for the training and test of SVM greatly determines the performance of the SVM system. In this paper, a new SVM speaker verification method based on adapted GMM-clustering feature was proposed, in which adapted GMM was used to extract a small quantity of typical feature vectors from large numbers of speech data (MFCC) for its excellent scalability. Experiments on text-independent speaker verification in NIST'04 lside-lside data showed significant improvement compared to the baseline GMM-UBM system.%支持向量机作为说话人建模方法用于与文本无关的话者确认研究时,如何提取适合SVM训练和测试的特征参数直接影响话者确认系统的性能和效率.根据高斯混合模型(GMM)聚类能力强的特点,提出一种基于自适应GMM聚类的说话人特征参数提取方法,通过白适应的GMM聚类将大样本、混叠严重的MFCC特征参数聚为小样本的、代表说话人个性特征的特征参数,并用于与文本无关的SVM话者确认.在NIST'04 1side-1side数据库上的实验表明了该方法的有效性.
展开▼