forensics_0.3B_xlsr_wild_deepfake_classifier

forensics_0.3B_xlsr_wild_deepfake_classifier

Built for audio you don't control. Most deepfake detectors are trained and tested on clean lab recordings, then fall apart on messy, real-world audio — phone recordings, compressed clips, unknown microphones, unknown everything. This model is the member of the Forensics family tuned specifically for that gap: an XLS-R + AASIST detector that leads the family on the In-the-Wild benchmark, the closest thing to "audio as it actually shows up in the real world."

• Backbone: facebook/wav2vec2-xls-r-300m (~300M params, fully unfrozen)
• Pooling: AASIST graph-attention

Part of the Forensics family

Full family: huggingface.co/collections/eliya/forensics-speech-deepfake-detection-family

Training

Trained using an agentic training loop — see eliyasegev/autotrain.

Weighted cross-entropy loss, Adam optimizer, large effective batch size, 9 synthetic fakes sampled per real utterance, backbone unfrozen from the very first step.

Trained on 5M+ speech samples spanning a mix of proprietary and publicly available data — some of our own self-created data has been open-sourced as well. Built for robustness against a wide range of TTS generation methods — including some of the most advanced synthesis techniques available today — and tuned specifically for generalization to real-world, out-of-domain audio. Same heavy augmentation stack as the rest of the family — codec transcoding, noise augmentation, RIR, RawBoost, SpecAugment, FreqMask, splice/mix, and cross-class splice.

Results

1.03% EER on In-the-Wild — the best real-world generalization result in the Forensics family.

Files in this repo

Setup

pip install -r requirements.txt   # torch, torchaudio, transformers, safetensors
hf download eliya/forensics_0.3B_xlsr_wild_deepfake_classifier --local-dir .

Run

python inference.py <audio.wav>

(Optionally override the checkpoint: python inference.py <audio.wav> <checkpoint.pt>.)

Audio is auto-converted to mono / 16 kHz and trimmed/padded to 5 s.

Output

fake_probability: <0..1>      # threshold is domain-dependent — adjust to your use case; ~0.1-0.2 is usually the best range
bonafide_score:   <0..1>      # raw P(real)
verdict: REAL | FAKE

Example

$ python inference.py real_human.wav
fake_probability: 0.1102
bonafide_score:   0.8898
verdict: REAL

$ python inference.py tts_fake.wav
fake_probability: 0.7366
bonafide_score:   0.2634
verdict: FAKE

Higher fake_probability = more likely a deepfake. Score is 1 − sigmoid(logit), since the classifier is trained with label 1 = real, 0 = fake.

References

• XLS-R: Babu et al., 2021, "XLS-R: Self-Supervised Cross-Lingual Speech Representation Learning at Scale", arXiv:2111.09296
• AASIST: Jung et al., 2021 (ICASSP 2022), "AASIST: Audio Anti-Spoofing Using Integrated Spectro-Temporal Graph Attention Networks", arXiv:2110.01200
• RawBoost augmentation: Tak et al., 2021 (ICASSP 2022), arXiv:2111.04433
• FreqMask augmentation: Xie et al., 2024, arXiv:2408.06922
• SpeechFake dataset: ACL 2025, "SpeechFake: A Large-Scale Multilingual Speech Deepfake Dataset Incorporating Cutting-Edge Generation Methods", arXiv:2507.21463 — this training recipe reproduces its Table 6

License

CC-BY-NC-4.0 — free for personal and research use.

Citation

If you use this in research, evaluation, or training, please acknowledge it:

@misc{segev2026forensicsxlsr,
  author       = {Segev, Eliya},
  title        = {Forensics 0.3B XLS-R: Speech Deepfake Detection for Real-World Audio},
  year         = {2026}
}