基于GMM-UBM/SVM的维吾尔语电话语音监控系统

摘要

Abstract In this paper we discuss a telephone speech monitoring system based on GMM-UBM/SVM. GMM is a method commonly used in speaker recognition system. But its accuracy of recognition will be impacted because of the voice monitored is short in time, and the telephone-internet terminals as well as transmission lines have big background noise. In our research we put into full play the robustness of GMM and the advantage of SVM in its perfect discrimination on small quantity static data in designing the telephone speech monitoring system and have discussed the performance of the system with Uighur. For comparison, we also discussed the recognitions with vector quantisation ( VQ), weighted VQ (WVQ) and base line. The proposed method improves the recognition rate to 20. 2% higher than VQ' s and 16. 7% higher than WVQ' s respectively for every 10 seconds test speech out of 20 seconds voice in a training set of 50 target speakers.%讨论基于GMM-UBM/SVM的电话语音监控系统.GMM是说话人识别系统中使用的常用方式.但由于监控语音发话时间短暂,电话-互联网终端及传输线背景噪音大等因素影响了GMM的识别精度.基于GMM的鲁棒性及SVM对小量静态数据具有高分类的优势设计电话语音监控系统并通过维吾尔语研讨了系统性能.为了便于比较,同时也讨论了量化距离(VQ)、加权量化距离(WVQ)及基线系统的识别.在50个目标人训练集,每人发话时间为20秒时,对10秒测试语音提案方法识别率对比于VQ和WVQ法分别提高了20.2％及16.7％.