论文部分内容阅读
采用多系统融合可以降低话者确认系统的等误识率(EER),融合策略一般是将多个系统的得分进行线性加权.这种方式是将每个系统得分情况作统一考虑,没有关注到各个系统自身得分的实际分布,某个系统表现不好的测试得分可能会劣化其他系统的优秀得分,从而对整个系统性能的提升有限.提出一种基于置信区间的融合策略,对I-Vector基线系统和基频辅助系统各设置一个置信区间,只有当基线系统的得分不在置信区间内并且辅助系统得分处于其置信区间时,才将两个系统的得分进行融合,否则仍然采取基线系统的得分.本文的方法充分利用了各系统的得分置信度,在NIST数据库上的实验表明,相对IVector基线系统,融合系统的性能提升了12.37%.
Multi-system integration can reduce the error-of-recognition rate (EER) of the speaker verification system, and the fusion strategy generally weights the scores of multiple systems linearly. This method takes the scoring of each system into account and does not pay attention to The actual distribution of each system’s own score, the poor performance of a system test scores may degrade the excellent score of other systems, thus enhancing the overall system performance is limited.This paper presents a confidence interval based fusion strategy, I-Vector baseline The system and the AF subsystem each set a confidence interval, and the scores of the two systems are merged only when the baseline system scores are not within the confidence interval and the assistant system scores are in their confidence intervals, otherwise the baseline system scores are still taken. The method in this paper takes full advantage of the confidence score of each system. Experiments on the NIST database show that the performance of the fusion system is improved by 12.37% over the IVector baseline system.