The NEC-TT 2018 speaker verification system

This paper describes the NEC-TT speaker verification system for the 2018 NIST speaker recognition evaluation (SRE'18). We present the details of data partitioning, x-vector speaker embedding, data augmentation, speaker diarization, and domain adaptation techniques used in NEC-TT SRE'18 speaker verification system. For the speaker embedding front-end, we found that the amount and diversity of training data are essential to improve the robustness of the x-vector extractor. This was achieved with data augmentation and mixed-bandwidth training in our submission. For the multi-speaker test scenario, we show that x-vector based speaker diarization is promising and holds potential for future research. For the scoring back-end, we used two variants of probabilistic linear discriminant analysis (PLDA), namely, the Gaussian PLDA and heavy-tailed PLDA. We show that correlation alignment (CORAL) and CORAL+ unsupervised PLDA adaptation are effective to deal with domain mismatch.