联合因子分析中的本征信道空间拼接方法

为了使联合因子分析适用于多种信道条件下的文本无关说话人识别,提出了一种本征信道空间的正交拼接法.在多信道条件下,可以通过混合数据法或简单拼接法估计本征信道空间,但前者存在空间掩盖,后者虽解决了空间掩盖但引入了空间重叠.本文首先证明说话人建模和测试的核心运算是斜投影,基于上述证明,通过将待拼接空间正交的方法移除了空间重叠.在NIST SRE 2008核心评测数据库上的实验表明,本文所提算法优于混合数据法和简单拼接法.

Eigenchannel Space Combination Method of Joint Factor Analysis

For application of joint factor analysis on the condition of multiple channels in text-independent speaker recognition, this paper proposes an eigenchannel space orthogonal combination method. The eigenchannel space can be estimated by a mix data method or a simple combination method on the condition of multiple channels. However, the former has space masking effects while the latter introduces space overlapping effects. This paper proves that the core computation of the speaker enrollment and test is an oblique projection. Space overlapping effects can be removed subsequently by an orthogonal method based on the above proof. On the NIST SRE 2008 core tasks corpus, the proposed method has a better performance than the mix data method and the simple combination method.