摘要
针对现有i-vector说话人确认系统在测试语音为短语音时性能下降的问题,对短语音i-vector估计的不确定性进行分析,改进了i-vector提取中Baum-Welch统计量的计算.该方法利用赋予权重的历史测试信息以及通用背景模型中的参数信息来增加用于短语音Baum-Welch统计量计算的说话人个性信息.将改进统计量用于i-vector提取,针对不同时长短语音的实验表明,新系统的性能优于当前i-vector系统,等错误率(EER)和检测代价函数最小值(min DCF)分别下降了13~19%和8~23%.
Aiming at the problem of the performance degradation of the existing i-vector system in the short utterance speaker verification task,an improved Baum-Welch statistic is proposed by analyzing the source of the i-vector estimation uncertainty. The pre-estimated background model parameter information as well as the weighted historical test speech information encountered by the system is included in improved Baum-Welch statistic. The improved statistic is applied to the extraction of the current test speech i-vector. Experiments on different duration test speech show that the performance of the improved i-vector based system is superior to the existing i-vector system,such as the equal error rate( EER) and the minimum detection cost function( min DCF) decreased by 13 ~ 19% and 8 ~ 23%,respectively.
作者
王铮
傅山
WANG Zheng;FU Shan(School of Electronic Information and Electrical Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)
出处
《小型微型计算机系统》
CSCD
北大核心
2019年第11期2264-2268,共5页
Journal of Chinese Computer Systems
基金
国家电网公司华东分部科技项目(SA0301503)资助
关键词
说话人确认
短语音
高斯混合模型
身份向量
模型自适应
speaker verification
short utterance
Gaussian mixture model
i-vector
model adaptation
