AWaveFormer: Audio Wavelet Transformer Network for Generalized Audio Deepfake Detection

Rapid advancements in speech synthesis technology have made it easier to produce realistic synthetic speech, which poses serious threats to public privacy and security. Recent studies have investigated pre-trained models for feature extraction and adopted advanced architectures, including convolutional and graph neural networks, for deepfake detection. Although these methods improve detection performance to some extent, their generalization and robustness still face challenges. To address this issue, in this paper, we propose a forgery detection method based on the fusion of dual pre-trained features and an optimized Transformer architecture. Specifically, we use a cross-attention mechanism to fuse the audio features extracted from pre-trained Wav2vec 2.0 and WavLM, and replace the token mixer in the traditional Transformer architecture with wavelet transform and multi-scale pooling operations. By combining dual pre-trained features, we extract more comprehensive and discriminative features while optimizing the Transformer architecture, which not only reduces time complexity but also enhances detection performance. We conducted experiments on several datasets, and the results show that our model achieves impressive detection performance with EERs of 0.13%, 2.33%, 3.63%, 10.25%, 5.15%, and 0.21% on the ASVspoof 2019LA, ASVspoof 2021LA, ASVspoof 2021DF, In-the-Wild, Fake-or-Real, and ASVspoof 2015LA evaluation datasets, respectively.