采用支持向量机的说话者确认中的样本平衡

支持向量机在与文本无关的话者确认系统中已经取得了广泛的应用,但是在实际应用系统中获得的目标说话人样本与冒认者样本数量比一般在几千分之一,因此存在很严重的样本非平衡问题,冒认者样本选择的好坏直接影响到整个系统的性能。本文提出了两种挑选冒认者样本的方法。实验证明这些方法能有效地解决上述问题,性能比随机挑选冒认者样本的方法有了提升,经过在2004年NIST说话人识别数据库上进行测试,等错误率由9.3%降低到6.8%,错误率相对下降了26.9%。

SVM Based Training Data Balance for Speaker Verification

Support Vector Machine(SVM)has been widely used in text-independent speaker verification systems.However,there are lots of training data unbalance problems with this algorithm due to the insufficiency of the data from target speakers.These problems may introduce severe performance degradation in application.So,it will influence the entire system's performance directly by choosing the right impostor.In this paper,we propose two strategies to select impostor's samples in SVM training.Experiments show that the methods proposed in the paper can efficiently solve the above unbalance problems,and significantly improve the performance of the system compared with the traditional methods which are based on random data selection algorithms.By testing the methods in the NIST 2004 benchmark,we have significantly reduced the equal error ratio of the speaker verification system from 0.093 to 0.068.