ProSDD-ASV5: Speech Anti-Spoofing Arena wrapper

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

ProSDD (Mahapatra et al., 2026): a fully fine-tuned wav2vec 2.0 XLS-R 300M encoder, first trained (Stage I, bona fide LibriSpeech train-clean-100 + dev only) to predict masked frame-level speaker + prosody targets, then (Stage II) jointly fine-tuned with that objective and a binary real/fake head on ASVspoof 2024 (ASVspoof 5) train/dev (RawBoost series 3). Inference uses only the clean path: XLS-R CNN front-end + feature projection, latent sequence fixed at 200 frames (zero-padded), 24-layer transformer encoder, mean pool, Linear(1024,512)-ReLU-Linear(512,2). Input as the authors' data_utils_eval.py: 16 kHz, no amplitude normalisation, fixed 4 s (64000 samples) centre crop, shorter clips centre zero-padded. FP32; score = logits[:,1] (bona fide class). Authors' released checkpoint; no upstream license, so the weights are linked from the authors' Google Drive, not re-hosted.

• Weights: not re-hosted (the authors grant no license for them). Get the authors' checkpoint; its sha256 is in the meta.yaml notes.
• Params: 315.963522 M. Trained on LibriSpeech (stage 1) and ASVspoof 5 train/dev (stage 2).
• 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; FP16 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/.

Citation

@article{mahapatra2026prosdd,
  title   = {ProSDD: Learning Prosodic Representations for Speech Deepfake Detection against Expressive and Emotional Attacks},
  author  = {Mahapatra, Aurosweta and Ulgen, Ismail Rasim and Lee, Kong Aik and Andrews, Nicholas and Sisman, Berrak},
  journal = {arXiv preprint arXiv:2604.13229},
  year    = {2026}
}