DPDA-GA-TCM — Speech Anti-Spoofing Arena wrapper

Truong et al., ICASSP 2026 — DPDA-GA: dual-path data-augmented training with gradient alignment (PCGrad). Three back-ends released; this card is one of them (DPDA-GA-AASIST, DPDA-GA-TCM, DPDA-GA-Mamba).

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

DPDA-GA (dual-path data-augmented training with gradient alignment) applied to an XLS-R 300M + TCM-Conformer detector: a fully fine-tuned wav2vec 2.0 XLS-R 300M front-end projected to 144-d (BatchNorm + SELU), followed by 4 Conformer blocks with Head-Token (TCM) multi-head attention (4 heads, depthwise conv kernel 31) and a class token feeding a linear 2-way head. Trained on ASVspoof2019 LA with each batch seen both clean and RawBoost-augmented, the two losses' gradients aligned with PCGrad. Checkpoint = weight average of the authors' n-best checkpoints, reproducing their main.py averaging exactly: the release's 5th checkpoint (best_4.pth) is truncated on HuggingFace and cannot be loaded, so 4 checkpoints are averaged instead of 5, and main.py's averaging code drops best_0 and double-counts best_1, giving (2*best_1 + best_2 + best_3)/4. FP32, deterministic first-66800-sample window (shorter clips repeat-tiled).

• Checkpoint (commit-pinned): dpda_ga_tcm_avg4.pth
• Params: 321.2132 M. Trained on ASVspoof2019 LA train only.
• Score: logits[:, 1] (higher = bona fide). Input: first 66,800 samples @16 kHz, shorter clips repeat-padded; FP32. 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

@inproceedings{truong2026dpdaga,
  author    = {Truong, Duc-Tuan and Liu, Tianchi and Li, Junjie and Tao, Ruijie and Lee, Kong Aik and Chng, Eng Siong},
  title     = {Addressing Gradient Misalignment in Data-Augmented Training for Robust Speech Deepfake Detection},
  booktitle = {ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},
  year      = {2026},
  pages     = {16252--16256},
  doi       = {10.1109/ICASSP55912.2026.11464739}
}