论文部分内容阅读
针对基于因子分析模型的说话人确认系统评分的复杂性以及需要较大运算量的问题,文章直接利用话者因子的余弦距离相似度来计算评分。首先在训练阶段和测试阶段分别用因子分析的方法从语音中估计出话者因子,然后直接利用话者因子评分。对比SVM和其它的JFA-GMM-UBM话者确认系统,本文中所采用的系统训练阶段和测试阶段的流程相同,并且目标话者模型只需要存储话者因子,存储量少。在NIST2008数据库上的实验结果表明,余弦距离评分对比其它因子分析模型的评分方法,更加简单,并且话者确认系统的性能也有提高。
In view of the complexity of system scoring based on the factor analysis model and the problem of requiring a large amount of computation, the article directly calculates the score by the cosine distance similarity of the speaker factor. First of all, in the training phase and the testing phase, factor analysis is used to estimate the speaker factor from speech and then the speaker factor score is used directly. Compared with SVM and other JFA-GMM-UBM speaker verification systems, the system training phase and the test phase flow are the same in this paper, and the target speaker model only needs to store the speaker factor and has less memory. The experimental results on the NIST2008 database show that the cosine distance score is more simple compared to other factor analysis models, and the performance of the speaker confirming the system is also improved.