采用非监督得分规整和因子分析的说话人确认

Speaker Verification Based on Unsupervised Normalization and Factor Analysis

在文本无关的说话人确认中,规整算法能够有效地调整测试得分的分布.另外,利用前面已经得到的测试语句的得分来调整规整的参数可以取得更好的效果,这种规整叫做非监督得分规整.在本文中,借用开发集得分来建立说话人和冒认者得分的两个先验高斯分布函数,在实际的测试中,利用最大后验概率准则来对规整的模型参数进行调整.在采用因子分析的情况下,在NIST2006说话人识别测试1conv4w-1conv4w数据库上,能够取得等错误率5.26%.

In the text-independent speaker verification, the normalization algorithm can adjust the score dislribution. The pre- vious test scores can be used to update the parameters of the normalization, which is defined as unsupervised score normalization in this paper. The scores distributions of the target and impostor in the development corpus are set up as a prior, and the parameters of normalization are updated using the maximum a posterior（MAP）algorithm in each test process. In the NIST 2006 speaker recognition evaluation（SRE）1 conv4w-1 conv4w corpus, the equal error rate（EER）of the system based on the factor analysis and unsupervised score normalization is 5.26 %.