XLSR-Conformer-TCM-DF: Speech Anti-Spoofing Arena wrapper

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

XLSR-Conformer + TCM (Truong et al., Interspeech 2024), DF-track model: a fully fine-tuned wav2vec 2.0 XLS-R 300M front-end projected to 144-d (Linear + BatchNorm + SELU), followed by 4 Conformer blocks whose multi-head self-attention carries the Temporal-Channel Modeling (TCM) module (per-head channel/head tokens, classification-token enrichment with the mean temporal and mean head tokens; 4 heads, depthwise conv kernel 31) and a class token feeding a linear 2-way head. Trained on ASVspoof2019 LA train with RawBoost (the paper's DF-track setting is stationary signal-independent additive noise). Checkpoint = the authors' released fixed-length DF model avg_5_best.pth (their own main.py average of the 5 best-validation checkpoints), unchanged. FP32, deterministic first-66800-sample window (shorter clips repeat-tiled).

• Checkpoint (commit-pinned): model.safetensors
• Params: 321.2132 M. Trained on ASVspoof 2019 LA train (RawBoost, DF-track setting).
• 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/.

Citation

@inproceedings{truong24b_interspeech,
  title     = {{Temporal-Channel Modeling in Multi-head Self-Attention for Synthetic Speech Detection}},
  author    = {Truong, Duc-Tuan and Tao, Ruijie and Nguyen, Tuan and Luong, Hieu-Thi and Lee, Kong Aik and Chng, Eng Siong},
  booktitle = {Interspeech 2024},
  pages     = {537--541},
  year      = {2024},
  doi       = {10.21437/Interspeech.2024-659}
}