Improving the Generalized Performance of Deep Embedding for Text-Independent Speaker Verification

In this paper, we propose an effective approach to improve the generalized performance of the speaker embedding system based on a neural network. The deep embedding system, based on x-vector, can obtain a good discriminative capability between different speakers while in the training stage. However, the cross-entropy loss function and the one-hot label limit the generalization ability of this kind of systems. This paper proposes a new training architecture by adding one filler node in the output layer to represent the out-of-domain speakers. Moreover, we adopt the attention mechanism in the proposed generalized x-vector system. From the results on the NIST 2010 SRE, we find that the proposed system achieves improved performance for the evaluation of varying duration conditions. On the full length condition, our proposed system obtains a 41.5% and 12.1% relative improvement over the baseline system (i-vector) and the standard x-vector system in equal error rate (EER).