DeepfakeDetector: audio models
Two models that classify speech as bona fide (REAL) or spoofed (FAKE). They are the audio detector of DeepfakeDetector, a diploma project.
| File | Use |
|---|---|
w2v2/w2v2.onnx, w2v2/w2v2.onnx.data |
Fine-tuned facebook/wav2vec2-xls-r-300m with a classification head ("accurate" mode); ONNX with external weights |
mel_resnet/mel_model-epoch=49-val_eer=0.0816.ckpt |
CNN on log-mel spectrograms ("fast" mode and spectrogram heatmaps), PyTorch Lightning checkpoint |
Preprocessing, thresholds and the inference code are in audio-detector/ of the GitHub repository. The detector
downloads the files from a pinned commit of this repository and checks their SHA-256 (audio-detector/models.toml).
Training data and licence
Trained on ASVspoof 2019 LA. The weights are released for research and education only, not for commercial use, following the terms of the training data. The wav2vec 2.0 XLS-R base model is Apache-2.0.
Limitations
ASVspoof 2019 covers a limited set of TTS and voice conversion systems. Newer generators, music, noise and phone audio can lower accuracy. Treat the output as a signal, not proof.