Add DPDA-GA-TCM submission for LRLspoof (1-SRR=24.6081)

submissions/dpda-ga-tcm.yaml

@@ -0,0 +1,61 @@
+schema_version: 4
+system:
+  name: DPDA-GA-TCM
+  slug: dpda-ga-tcm
+  description: '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).'
+  code: https://github.com/ductuantruong/dpda_ga
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/DPDA-GA-TCM/blob/e949ad036fe984b80fadaa629ac5544f04e0ac26/dpda_ga_tcm_avg4.pth
+  params_millions: 321.2132
+  paper:
+    arxiv_id: '2509.20682'
+    url: https://arxiv.org/abs/2509.20682
+    bibtex: "@inproceedings{truong2026dpdaga,\n  author    = {Truong, Duc-Tuan and Liu, Tianchi and Li,\
+      \ Junjie and Tao, Ruijie and Lee, Kong Aik and Chng, Eng Siong},\n  title     = {Addressing Gradient\
+      \ Misalignment in Data-Augmented Training for Robust Speech Deepfake Detection},\n  booktitle =\
+      \ {ICASSP 2026 - 2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)},\n\
+      \  year      = {2026},\n  pages     = {16252--16256},\n  doi       = {10.1109/ICASSP55912.2026.11464739}\n\
+      }\n"
+dataset:
+  id: lab260/LRLspoof
+  revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
+  split: test
+scores:
+  srr_complement: 24.60808376828463
+  n_trials: 1304455
+  n_skipped: 286
+calibration:
+  source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
+  threshold: -0.9581514775141243
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/DPDA-GA-TCM/resolve/acbe282f7cde0cb486216009ea21aff7ad2f7b36/.eval_results/lab260/LRLspoof/scores.txt
+  scores_sha256: 20fb892b402fe41d2121d5b4d0a70234de8ffdc4dfd431a0a121dd53432aac90
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-11-01'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-01'
+notes: 'DPDA-GA XLS-R + TCM-Conformer (PCGrad) from ductuantruong/dpda_ga @ aab8961 (MIT), weights from
+  HF tuantd/dpda_ga @ 531554e pcgrad/conformer_tcm/best/best_{0..4}.pth. best_4.pth is truncated on HF
+  (547,680,256 of ~1,285,107,746 bytes; no ZIP central directory), so the authors'' 5-checkpoint average
+  (main.py --average_model True --n_average_model 5) cannot be rebuilt; we apply main.py''s averaging
+  to the 4 intact checkpoints. main.py''s loop has a state_dict aliasing bug (sd = model.state_dict()
+  shares storage with the model, so loading best_1 overwrites it): its result is (2*b1 + b2 + ... + b_{n-1})/n
+  with best_0 dropped. We reproduce it exactly: (2*best_1 + best_2 + best_3)/4 (make_avg.py). Strict load.
+  Deterministic first-66800-sample window, shorter clips repeat-tiled (source utils.pad), no resampling/normalisation.
+  score = logits[:, 1] (bona fide = class 1; source produce_evaluation_file); higher = more bona fide.
+  FP32 with cuDNN TF32 off (TF32 convs made scores batch-dependent by ~2e-4). Paper EERs (5-checkpoint
+  model): 21DF 1.81%, ITW 6.48%. Trained on ASVspoof2019 LA train (dev for model selection) only.'