A Study of Joint Framework for Robustness Against Noise on Speaker Verification System.

As speaker verification (SV) technology is applied to various fields with the development of deep learning, robustness against noise is becoming more prominent. A common approach for suppressing the noise effect is to jointly train SV system along with a front-end enhancement module as pre-processing. However, in this paper, we explore a training method for a noise-robust system that focuses on enhancing the single SV system. In particular, we introduce the feature-robust loss to exploit the embeddings extracted from the pre-trained model, which is jointly trained with the enhancement module. Experimental results showed that our proposed method can mitigate the noise effect in various noisy conditions using only a single SV system, demonstrating its potential for further development.