On the use of X-vectors for Robust Speaker Recognition

Text-independent speaker verification (SV) is currently in theprocess of embracing DNN modeling in every stage of SV system.Slowly, the DNN-based approaches such as end-to-endmodelling and systems based on DNN embeddings start to becompetitive even in challenging and diverse channel conditionsof recent NIST SREs. Domain adaptation and the need for alarge amount of training data are still a challenge for currentdiscriminative systems and (unlike with generative models), wesee significant gains from data augmentation, simulation andother techniques designed to overcome lack of training data.We present an analysis of a SV system based on DNN embeddings(x-vectors) and focus on robustness across diverse datadomains such as standard telephone and microphone conversations,both in clean, noisy and reverberant environments. Wealso evaluate the system on challenging far-field data createdby re-transmitting a subset of NIST SRE 2008 and 2010 microphoneinterviews. We compare our results with the stateof-the-art i-vector system. In general, we were able to achievebetter performance with the DNN-based systems, but most importantly,we have confirmed the robustness of such systemsacross multiple data domains.