Spoofing-Aware Speaker Verification Robust Against Domain and Channel Mismatches

Abstract

Building speaker verification systems that are simultaneously robust to spoofing attacks, channel mismatch, and domain mismatch remains difficult in real-world use. Traditional systems often handle these issues independently, leading to degraded performance under combined conditions. We propose an integrated framework that incorporates pair-wise learning and spoofing attack simulation into meta-learning, using an asymmetric dual-path model with multi-task training for ASV, anti-spoofing, and spoofing-aware ASV. We also introduce CNComplex, a testing dataset designed to evaluate these joint threats. Experiments show substantial gains over conventional ASV systems, demonstrating strong robustness and generalization in practical scenarios.