Audio Anti-Spoofing Based on Audio Feature Fusion

The rapid development of speech synthesis technology has significantly improved the naturalness and human-likeness synthetic speech. As technical barriers for are rapidly lowering, number illegal activities such as fraud extortion is increasing, posing a significant threat to authentication systems, automatic speaker verification. This paper proposes an end-to-end detection model based on audio feature fusion in response constantly evolving techniques improve accuracy detecting uses pre-trained wav2vec2 extract features from raw waveforms utilizes module back-end classification. aims by adequately utilizing extracted front end fusing information timeframes dimensions. Data augmentation also used enhance performance generalization model. trained training sets logical access (LA) dataset ASVspoof 2019 Challenge, international standard, tested deep-fake (DF) evaluation datasets 2021 Challenge. equal error rate (EER) LA DF 1.18% 2.62%, respectively, achieving best results dataset.