Post-training for Deepfake Speech Detection
About
We introduce a post-training approach that adapts self-supervised learning (SSL) models for deepfake speech detection by bridging the gap between general pre-training and domain-specific fine-tuning. We present AntiDeepfake models, a series of post-trained models developed using a large-scale multilingual speech dataset containing over 56,000 hours of genuine speech and 18,000 hours of speech with various artifacts in over one hundred languages. Experimental results show that the post-trained models already exhibit strong robustness and generalization to unseen deepfake speech. When they are further fine-tuned on the Deepfake-Eval-2024 dataset, these models consistently surpass existing state-of-the-art detectors that do not leverage post-training. Model checkpoints and source code are available online.
Related benchmarks
| Task | Dataset | Result | Rank | |
|---|---|---|---|---|
| Audio Deepfake Detection | in the wild | EER1.23 | 58 | |
| Audio Deepfake Detection | ITW | ACC98.7 | 15 | |
| Speech Deepfake Detection | FakeOrReal | EER173 | 9 | |
| Speech Deepfake Detection | ODSS | EER (%)1.13 | 7 | |
| Speech Deepfake Detection | EF | EER20 | 7 | |
| Speech Deepfake Detection | ADD ASVspoof 2022 | EER1.05 | 7 | |
| Speech Deepfake Detection | ADD ASVspoof 2023 | EER4.67 | 7 | |
| Speech Deepfake Detection | DV | EER2.27 | 7 | |
| Speech Deepfake Detection | FSW | EER (%)16.15 | 7 | |
| Speech Deepfake Detection | FoR | Accuracy98.05 | 6 |