A Study on Bias and Fairness In Deep Speaker Recognition

The protected With the ubiquity of smart devices that use speaker recognition (SR) groups in this study are defined as'gender' and'nationality' systems as a means of authenticating individuals and personalizing of speakers. The VoxCeleb dataset [4] is used in this study given their services, fairness of SR systems has becomes an important that it is the most widely used dataset in training SR systems and the point of focus. In this paper we study the notion of fairness in recent wide range of diversity among its speakers. SR systems based on 3 popular and relevant definitions, namely Statistical In summary we make the following contributions: (1) We evaluate Parity, Equalized Odds, and Equal Opportunity. We examine fairness of the current widely used architectures in SR and crossexamine 5 popular neural architectures and 5 commonly used loss functions them with different loss functions used in training.