Language Adaptation for Speaker Recognition Systems Using Contrastive Learning.

In this article we propose to study several approaches to adapt a system between two languages. To train the state of the art x-vector Speaker Verification system, we need a huge amount of labeled speech data. If this constraint is satisfied in English (due to Voxceleb), it is not in our target domain: French. We use a supervised Contrastive Learning to transfer knowledge between source and target domain. Among the two other proposed adaptation approaches (Multilingual Learning and Transfert Learning) we show that the one based on Contrastive Learning gives the best performance: about 30% relative gain in term of Equal Error Rate with respect to the baseline system. We also show the robustness of the Contrastive Learning with respect to the duration (from very short to short) as well as to distortion presence (noise, reverberation).