LAUSR

Abstract With the development of deep learning techniques, speaker verification (SV) systems based on deep neural network (DNN) achieve competitive performance compared with traditional i-vector-based works. Previous DNN-based SV methods usually employ time-delay neural network, limiting the extension of the network for an effective representation. Besides, existing attention mechanisms used in DNN-based SV systems are only applied to a single level of network architectures, leading to insufficiently extraction of important features. To address above issues, we propose an effective deep speaker embedding architecture for SV, which combines a residual connection of one-dimensional dilated convolutional layers, called dilated residual networks (DRNs), with a multi-level attention model. The DRNs can not only capture long time-frequency context information of features, but also exploit information from multiple layers of DNN. In addition, the multi-level attention model, which consists of two-dimensional convolutional block attention modules employed at the frame level and the vector-based attention utilized at the pooling layer, can emphasize important features at multiple levels of DNN. Experiments conducted on NIST SRE 2016 dataset show that the proposed architecture achieves a superior equal error rate (EER) of 7.094% and a better detection cost function (DCF16) of 0.552 compared with state-of-the-art methods. Furthermore, the ablation experiments demonstrate the effectiveness of dilated convolutions and the multi-level attention on SV tasks.