Pellav3: 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No paper: the model sits in the Arena's unpublished tier (shown, not ranked).
Pellav3 (Sadanie, November 2026): an open-weights two-view audio deepfake detector (1.54 B parameters). Each 4 s window is read by XLS-R 1B on the standardised waveform (softmax mix of all 49 hidden states) and by w2v-BERT 2.0 on 80-bin kaldi log-mel fbank features (25-state mix); each view is projected to 256-d and refined by two Res2 blocks, the frame sequences are concatenated per frame (512-d), refined by two shared Res2 blocks and pooled by attentive statistics pooling into Linear(1024,1) -> logit(fake). Authors' released checkpoint; training data undisclosed. Input as the authors' pellav3_infer.py view(): 16 kHz, mean removal, 20 Hz high-pass (zero-phase 2nd-order Butterworth), ITU-R BS.1770 loudness normalisation to -23 LUFS (pyloudnorm semantics), zero-pad to 4 s. Arena 4 s cap: one window, the first 4 s (64000 samples) of the authors' window grid (the authors average 4 s windows every 2 s over longer clips; clips <= 4 s are scored identically). score = -logit(fake), higher = bona fide. FP32 (TF32 off; the authors run bfloat16 autocast).
- Weights: re-hosted here under the upstream license (pellav3_detector.safetensors, commit-pinned); upstream: https://huggingface.co/Sadanie/pellav3-audio-deepfake-detector/blob/47b69420803c01de65af264ecdc2c5f35f2184f3/pellav3_infer.py
- Params: 1544.09 M. Trained on undisclosed data (no paper; the card is silent), so no Arena cell is marked n/a*, but leakage cannot be ruled out (J-SPAW_LA scores are perfectly separated).
- Score: higher = bona fide. J-SPAW_LA was scored with the PyTorch wrapper; most 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 (|dEER| <= 0.25, Spearman >= 0.999); the RTX 4060 Ti cards ran PyTorch.
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/.
- Downloads last month
- 213