Add LaST-Net submission for DECRO (EER=16.0787)

submissions/last-net.yaml

@@ -0,0 +1,55 @@
+schema_version: 4
+system:
+  name: LaST-Net
+  slug: last-net
+  description: 'LaST-Net (Length-Aware Layer and Scale-Adaptive Temporal Network; Jun Xue, 2026): a fine-tuned
+    fairseq XLS-R 300M front-end whose 24 transformer-layer outputs are fused by a length- and content-conditioned
+    layer router, projected to 192-d, processed at three temporal scales (1x, 1/2, 1/4) by a length-selective
+    kernel block and dilated depthwise residual blocks with two length-prior residual scale-exchange stages,
+    attentive statistics pooling per scale and a 2-way MLP head. The model also takes the clip''s available
+    duration (length features). Trained on ASVspoof 2019 LA train (dev for checkpoint selection) with
+    RawBoost-4 paired prediction consistency. Authors'' released epoch-52 checkpoint. Input exactly as
+    the authors'' infer.py at --seconds 4: 16 kHz, no normalisation, first 64000 samples (4 s), shorter
+    clips repeated (np.tile) and cropped, available seconds = min(len, 64000)/16000 passed to the model.
+    score = log_softmax(logits)[:, 1] (the authors'' bonafide_log_score). FP32 (TF32 off).'
+  code: https://github.com/JunXue-tech/LaST-Net/tree/3838f0c7108313f771f6c1e208ac8f1e3bde60b0
+  checkpoint: https://huggingface.co/JunXueTech/LaST-Net/blob/7b18ff0abff078c60745aaaaf5165e3f9542fa1b/best.pt
+  params_millions: 316.955664
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/DECRO
+  revision: e54c29df5b81bac75e5cc294d1ab0d4db72a5c71
+  split: test
+scores:
+  eer_percent: 16.078664634373023
+  n_trials: 37314
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/LaST-Net/resolve/c8c07b94446687a23d798ef5e1c465460a89b8b9/.eval_results/SpeechAntiSpoofingBenchmarks/DECRO/scores.txt
+  scores_sha256: b79c70ddf4dc8efc233e94379ba196b9e0331ded505ab6e4b2446f53f9010ddf
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-11-10'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-10'
+notes: 'LaST-Net = github.com/JunXue-tech/LaST-Net @ 3838f0c7108313f771f6c1e208ac8f1e3bde60b0 (MIT; model.py,
+  infer.py, eval.py, load_dataset.py), weights HF JunXueTech/LaST-Net @ 7b18ff0abff078c60745aaaaf5165e3f9542fa1b
+  best.pt (sha256 8fee6178d19b398fac9e1b1bfcf13720c4eb5ff722b5b52e4bc698e0de0fa28b; epoch 52, includes
+  optimizer + RNG state). No weight license -> linked, not re-hosted. best.pt was opened once with torch.load(weights_only=True)
+  plus an explicit numpy allow-list (RNG-state metadata) and its "model" dict (522 fp32 tensors) saved
+  as safetensors (sha256 3ec3ff415654931474fe0949b5ed1030a84b8c0b2ee28a4fce37f238e952b264); the wrapper
+  reads only that file. XLS-R 300M is built with fairseq 0.12.2 from the xlsr2_300m.pt model cfg and every
+  tensor is overwritten (strict load, 0 missing / 0 unexpected); 10 loaded XLS-R pre-training tensors
+  (mask_emb, quantizer, project_q, final_proj, encoder.layer_norm) are off the inference path (features_only,
+  layer=23). Preprocessing and score exactly as infer.py at 4 s (64000 samples, prefix crop / repeat,
+  available-duration input); verified against the authors'' modules (max |d| 1.6e-5 at 4 s, 1.8e-5 at
+  6 s, 19 J-SPAW_LA clips). Deviations: Arena 4 s window (the authors report 1-6 s; 4 s is one of their
+  settings and uses 64000, not 64600, samples); strict FP32 instead of the authors'' TF32 matmuls. Training:
+  ASVspoof 2019 LA train (dev for selection); the Arena ASVspoof sets are eval partitions -> no trained_on.
+  No paper found (arXiv / web search 2026-11-09) -> unpublished tier. Self-reported EER% at 4 s: 19LA
+  0.09 / 21LA 3.75 / 21DF 2.70 / ITW 6.47 (1-6 s averages 1.29 / 4.98 / 3.62 / 7.58). J-SPAW_LA EER 14.500%
+  (PyTorch; TensorRT FP32 identical on RTX 4070 Ti SUPER and RTX 5060 Ti, Spearman 1.0).'