The I2R’s Submission to VOiCES Distance Speaker Recognition Challenge 2019

This paper is about the I2R’s submission to the VOiCES from a distance speaker recognition challenge 2019. The submissions were based on the fusion of two x-vectors and two i-vectors subsystems. Main efforts have been focused on the frontend de-reverberation processing, PLDA backend design, score normalization and fusion studies in order to improve the system performance on single channel distant/far-field audio, under noisy conditions. We contribute to the fixed condition task under specific training and development data set. The experimental results showed that the de-reverberation approach can achieve 5% to 10% relative improvement on both EER and DCF for all subsystems and more than 10% improvement in the final fusion system on the Dev dataset and more than 15% relative improvement on the final evaluation dataset. Our final fusion system achieved about 2% EER rate and 0.240 minDCF on the Development Dataset.