论文部分内容阅读
对于采用统一阈值的,基于高斯混合模型(GMM)的文本无关说话人确认系统,由于不同的话者模型的输出评分分布的不同,会影响到系统的确认性能,为此,需对输出评分进行规整。本文提出了一种新的评分规整方法-整体规整。整体规整同时考虑了不同测试语音和不同话者模型的差异,并在评分域做出调整,使得所有语音的输出评分具有相似的分布,从而使系统整体分类能力得以保证。在NIST’03电话语音库上进行的实验表明,采用了整体规整后的系统性能和传统的评分规整方法比较,有了明显提高。
For a GMM-based text-independent speaker verification system that uses a uniform threshold, the output rating will be adjusted due to the difference in the output score distribution of different speaker models, which will affect the confirmation performance of the system . In this paper, we propose a new scoring method - integral regularity. The whole regulation also takes into account the differences between different test voices and different speaker models and makes adjustments in the scoring field to make the output scores of all speech have similar distributions so that the overall system classification ability can be guaranteed. Experiments conducted on the NIST ’03 telephone voice library show that the system performance using the whole system has been significantly improved over the traditional scoring system.