Abstract: This paper presents a method of combining Deep Neural Network（DNN） regression model with i-vector/Probabilistic Linear Discriminant Analysis（PLDA） speaker recognition system. By fitting the nonlinear function relationship between the noisy and the clean speech i-vector using DNN, the approximate representation of the clean speech i-vector can be obtained to achieve the purpose of reducing the influence of the noise on the system performance. The feasibility and effectiveness of the proposed method are verified by the experiments on TIMIT data set.

Key words: speaker recognition, Deep Neural Network（DNN）, i-vector

摘要： 提出了一种将基于深度神经网络（Deep Neural Network，DNN）特征映射的回归分析模型应用到身份认证矢量（identity vector，i-vector）/概率线性判别分析（Probabilistic Linear Discriminant Analysis，PLDA）说话人系统模型中的方法。DNN通过拟合含噪语音和纯净语音i-vector之间的非线性函数关系，得到纯净语音i-vector的近似表征，达到降低噪声对系统性能影响的目的。在TIMIT数据集上的实验验证了该方法的可行性和有效性。

关键词: 说话人识别, 深度神经网络, i-vector