eliya / forensics_0.3B_wavlm_oc_softmax_deepfake_classifier

forensics_0.3B_wavlm_oc_softmax_deepfake_classifier

WavLM + AASIST deepfake speech detector trained with a combined CE + OC-Softmax objective, which pulls bonafide embeddings into a tight cluster and pushes every spoof type outside it — a different decision-boundary shape than the plain-CE model, useful as an ensemble partner.

Detects whether a speech clip is real (bonafide) or a deepfake (spoof) .

• Backbone: microsoft/wavlm-large (~300M params)
• Pooling: AASIST graph-attention (gat_dims 128/64)
• Training run: wavlm_aasist_27_03_oc_softmax_ce_from_16 (last epoch, #10 of this sub-run, shipped)

Part of the Forensics family

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

Training (inferred)

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

CE + OC-Softmax combined loss, continuing from the epoch-16 checkpoint of wavlm_aasist_27_03_oc_softmax . Same core data sources as the other models in this batch (SpeechFake, DFADD, CodecFake, MD-CommonVoice, ASVspoof2019-LA) and the same augmentation stack (codec, MUSAN, RIR, RawBoost, SpecAugment, splice/mix).

Results — epoch 10 / "last" (shipped checkpoint)

Note: "last" was shipped rather than the metric-best epoch — epoch 8 has a slightly lower avg-all (3.66% vs 3.73%), though epoch 10 has a better In-the-Wild EER (1.76% vs 1.92%). It's a close call either way; worth a quick look before assuming epoch 10 is optimal.

Files in this repo

Setup

pip install -r requirements.txt   # torch, torchaudio, transformers, safetensors
hf download eliya/forensics_0.3B_wavlm_oc_softmax_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: 0.0984
bonafide_score:   0.9016
verdict: REAL

$ python inference.py tts_fake.wav
fake_probability: 0.8317
bonafide_score:   0.1683
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

• WavLM: Chen et al., 2022, "WavLM: Large-Scale Self-Supervised Pre-Training for Full Stack Speech Processing" , arXiv:2110.13900
• AASIST: Jung et al., 2021 (ICASSP 2022), "AASIST: Audio Anti-Spoofing Using Integrated Spectro-Temporal Graph Attention Networks" , arXiv:2110.01200
• OC-Softmax: Zhang et al., 2020 (IEEE SPL 2021), "One-Class Learning Towards Synthetic Voice Spoofing Detection" , arXiv:2010.13995
• RawBoost augmentation: Tak et al., 2021 (ICASSP 2022), arXiv:2111.04433
• SpeechFake dataset: ACL 2025, "SpeechFake: A Large-Scale Multilingual Speech Deepfake Dataset Incorporating Cutting-Edge Generation Methods" , arXiv:2507.21463

License

CC-BY-NC-4.0 — free for personal and research use. For commercial use, contact [email protected] .