Vocoded-VoxCeleb2-SSL-P1: Speech Anti-Spoofing Arena wrapper

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

NII vocoded-VoxCeleb2 SSL countermeasure, model ID P1 of Wang & Yamagishi (ICASSP 2024): a single wav2vec 2.0 front end, "v.vox": XLSR-53 continually pre-trained (self-supervised, fairseq) on >9,000 h of VoxCeleb2 copy-synthesised by multiple neural vocoders. Back end: Linear 1024->128, three Linear+LeakyReLU+MC-dropout (p=0.5) blocks, mean pooling over frames and Linear 128->2. Whole model fine-tuned on voc.LA (vocoded ASVspoof 2019 LA train bona fide, ~10 h) with RawBoost. Score = logit of the mean bona-fide softmax over 144 Monte-Carlo-dropout samples (the authors' scoring), with fixed per-clip dropout masks. Authors' released seed-01 checkpoint (BSD-3). FP32, deterministic first-64600-sample window (shorter clips repeat-tiled) -- NOT the paper's full-utterance protocol.

• Checkpoint (commit-pinned): model.safetensors
• Params: 317.5716 M. Trained on vocoded ASVspoof 2019 LA train bona fide (voc.LA); SSL front end continually pre-trained on vocoded VoxCeleb2.
• 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; PyTorch on the RTX 5060 Ti (no engine passed the parity bar there)), 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

@inproceedings{wang2024vocoded,
  title     = {Can Large-Scale Vocoded Spoofed Data Improve Speech Spoofing Countermeasure with a Self-Supervised Front End?},
  author    = {Wang, Xin and Yamagishi, Junichi},
  booktitle = {ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year      = {2024},
  eprint    = {2309.06014},
  archivePrefix = {arXiv}
}