Language-adversarial training for indic multilingual speaker verification

Abstract

Speaker verification now reports a reasonable level of accuracy in its applications in voice-based biometric systems. Recent research on deep neural networks and predicting speaker identity based on speaker embeddings has gained remarkable success. However, results are limited when it comes to verifying multilingual speakers. In this paper, we propose an ensemble system submitted to the I-MSV Challenge 2022. The system is built upon the ECAPA-TDNN and RawNet2 models with additional adversarial training layers. Probabilistic Linear Discriminant Analysis (PLDA) back-end scoring and Large Margin Cosine Loss (LMCL) are implemented to further obtain more discriminative features. Experimental results show that on the Constraint Private Test set of the task, our proposed model achieved remarkable results, ranking third with an Equal Error Rate (EER) of 2.9734%.