在说话人确认系统的实际应用中,让用户提供大量的训练语音是不现实的,所以在GMM-SVM系统中,正样本点数通常只有一个,而负样本点数远远多于正样本点数,造成SVM分类超平面严重偏向负样本,这种情况对于支持向量机的性能影响很大。针对此问题,提出了基于时间间隔对语音数据进行分段的方法,来增多正样本点数,得到更好的分类超平面。美国国家标准与技术研究所(NIST)2002年说话人识别数据库上的实验证明,语音分段的方法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。 In the practical application of speaker recognition system, it is unrealistic to let users provide a large amount of training speech. Therefore, in GMM-SVM system, there is usually only one positive sample point while negative sample points are far more than positive sample points, resulting in SVM classification Hyperplanes are heavily biased toward negative samples, which has a great impact on the performance of SVM. To solve this problem, a method of segmenting speech data based on time interval is proposed to increase the number of positive samples and obtain a better classification hyperplane. Experiments on the 2002 National Institute of Standards and Technology (NIST) speaker identification database show that the speech segmentation method can improve the recognition accuracy and robustness of the entire speaker verification system to a certain extent.