ASV-CM Score Fusion for Robust Speaker Verification in Spoofed Scenarios

• Lucas A. R. Carvalho ICTi
• Laiza P. Lima ICTi
• Guilherme Rinaldo ICTi
• Cinthya O. Silva ICTi
• Victor C. Beraldo ICTi

Resumo

Spoofing-Aware Speaker Verification (SASV) hybrid systems have become a necessity as traditional Automatic Speaker Verification (ASV), despite recent improvements, remain vulnerable to spoofing attacks. This work proposed a multi-model fusion that integrates ASV and Countermeasure (CM) systems by combining decision scores from both types. The individual and combined systems were evaluated on their ability to detect spoofing. Results show that spoofing attacks significantly degrade the performance of all standalone ASV systems. Among the CM models, Whisper and HuBERT outperformed the ASVspoof2019 baselines on both the Logical Access (LA) and Physical Access (PA) datasets. Among all hybrid configurations evaluated, the WeSpeaker Resnet34-HuBERT and WeSpeaker Resnet34-Whisper combinations demonstrated the best performance, achieving EER (Equal Error Rate) of 0.5% and 0.8% on the LA and PA datasets, respectively. These results achieved performance close to the state-of-the-art while offering improved and better generalization.