Forensics-WavLM-Base: Speech Anti-Spoofing Arena wrapper

Packaged for the Speech Anti-Spoofing Arena from the authors' code.

No paper: the model sits in the Arena's unpublished tier (shown, not ranked).

Forensics 0.3B Base speech deepfake detector (Eliya Segev, 2026): a fully fine-tuned WavLM-Large front-end whose 24 transformer hidden states are pooled by an AASIST-style spectro-temporal graph-attention pool (layer x time graphs, GAT dims 128/64), followed by a 3-block residual MLP classifier (320-d) and a single logit (label 1 = real); trained with label-smoothed CE + OC-Softmax + real-anchor supervised-contrastive losses. Authors' checkpoint_epoch_5 as released. Input as the authors' inference.py: peak-normalised waveform, 5 s (80000 samples) centre crop, shorter clips repeat-padded. score = the logit (sigmoid = P(real)). Unpublished model (HF model card only, no paper).

• Weights: re-hosted here under the upstream license (checkpoint_epoch_5.safetensors, commit-pinned); upstream: https://huggingface.co/eliya/forensics_0.3B_base_deepfake_classifier/tree/cfd11365b02167214dc3b392378809fef6ca3cfa
• Params: 317.257863 M. Trained on the author's undisclosed ~5M-clip mix incl. ASVspoof 2019 LA, ASVspoof 2021 DF, DFADD and In-the-Wild. The Arena marks ASVspoof2019_LA, ASVspoof2021_DF, DFADD and InTheWild as n/a* (training overlap).
• Score: higher = bona fide. J-SPAW_LA was scored with the PyTorch wrapper; the other datasets with the TensorRT drop-in class (FP32 on the RTX 4070 Ti SUPER; FP32 on the RTX 5060 Ti), which matched PyTorch on the full J-SPAW_LA set.

Arena results

EER% (lower is better); EmoSpoofTTS and LRLspoof are spoof-only and use 1-SRR% at the model's DeepVoice EER threshold. Per-dataset scores are under .eval_results/.