CAARMA introduces a novel approach to improve speaker verification systems by addressing data limitations through strategic data augmentation.

• Employs adversarial mixup regularization to create synthetic training examples
• Enhances model robustness by improving embedding space clustering
• Particularly valuable for scenarios with limited speaker data
• Strengthens biometric security applications by reducing false acceptance/rejection rates

This research significantly advances voice-based authentication systems by creating more resilient models that can better distinguish between speakers, even with limited training data. These improvements directly enhance security applications that rely on voice biometrics.

CAARMA: Class Augmentation with Adversarial Mixup Regularization