高斯PLDA在说话人确认中的应用及其联合估计

Gaussian PLDA for Speaker Verification and Joint Estimation

近年来,基于总变化因子的说话人识别方法成为说话人识别领域的主流方法.其中,概率线性鉴别分析(Probabilistic linear discriminant analysis,PLDA)因其优异的性能而得到学者们的广泛关注.然而,在估计PLDA模型时,传统的因子分析方法只更新模型空间,因此,模型均值不能很好地与更新后的模型空间耦合.提出联合估计法对模型均值和模型空间同时估计,得到更为严格的期望最大化更新公式,在美国国家标准与技术局说话人识别评测2010扩展测试数据库以及2012核心测试数据库上,等错率得到一定提升.

Recently the approaches based on i-vector have become very popular in the speaker recognition domain. Among these methods, the probabilistic linear discriminant analysis （PLDA） has attracted much attention due to its promising performance. However, the traditional factor analysis method only updates model space, thus making model mean couple with the model space unsuitably. This paper propose an approach of joint estimation for both model mean and model space, resulting in more strict expectation maximization （EM） formula. The equal error rate has been improved on the NIST SRE 2010 extended test corpus and NIST SRE 2012 core test corpus.