ELEAT-SAGA: Early & Late Integration with Evading Alternating Training for Spoof-Robust Speaker Verification

Researchers introduced ELEAT-SAGA, a novel architecture for spoof-robust speaker verification. It uses a 'score-aware gated attention' mechanism to dynamically adjust speaker embeddings based on anti-spoofing scores. The model integrates pre-trained ECAPA-TDNN and AASIST models with new alternating training strategies. On the ASVspoof 2019 benchmark, it achieved a state-of-the-art spoofing-aware speaker verification equal error rate (SASV-EER) of just 1.22%, significantly outperforming previous baselines.