Add AASIST3 submission for DFADD (#11)

- Add AASIST3 submission for DFADD (b00186364ead8679063349fb540b725d249d9ea8)

submissions/aasist3.yaml

@@ -0,0 +1,49 @@
+schema_version: 4
+system:
+  name: AASIST3
+  slug: aasist3
+  description: 'KAN-enhanced AASIST speech deepfake detector with a wav2vec 2.0 (XLS-R-53) self-supervised
+    front-end. SSL features are projected to 128-d through a Kolmogorov-Arnold (KAN) bridge, processed
+    by a RawNet2-style residual encoder and spectro-temporal graph-attention layers, and classified by
+    a four-branch inference head with a KAN output layer. ASVspoof 2024 Challenge system; the published
+    lab260/AASIST3 checkpoint (note: these weights differ from the paper results). FP32, preemphasis (0.97),
+    deterministic first-64600-sample window (no random crop). score = output logit for class 1 (bona fide).'
+  code: https://github.com/mtuciru/AASIST3
+  checkpoint: https://huggingface.co/lab260/AASIST3
+  params_millions: 321.7495
+  paper:
+    arxiv_id: '2408.17352'
+    url: https://arxiv.org/abs/2408.17352
+    bibtex: "@inproceedings{borodin24_asvspoof,\n  title={AASIST3: KAN-enhanced AASIST speech deepfake\
+      \ detection using SSL features and additional regularization for the ASVspoof 2024 Challenge},\n\
+      \  author={Borodin, Kirill and Kudryavtsev, Vasiliy and Korzh, Dmitrii and Efimenko, Alexey and\
+      \ Mkrtchian, Grach and Gorodnichev, Mikhail and Rogov, Oleg Y.},\n  booktitle={The Automatic Speaker\
+      \ Verification Spoofing Countermeasures Workshop (ASVspoof 2024)},\n  pages={48--55},\n  year={2024},\n\
+      \  doi={10.21437/ASVspoof.2024-8}\n}\n"
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/DFADD
+  revision: da9ec704a3db3b6a93b3f459fedd80e34b8ccdfa
+  split: test
+scores:
+  eer_percent: 1.4000000000000001
+  n_trials: 3755
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/lab260/AASIST3/resolve/9781bbc24551f969fb076e83c4823857b442761a/.eval_results/SpeechAntiSpoofingBenchmarks/DFADD/scores.txt
+  scores_sha256: 934715da01061eda91f0bd1e3bba56d5c59fde5d0f090c2978eb9e67fad608b9
+  bench_version: speech-spoof-bench==0.3.4
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-06-10'
+  reproduced_bench_version: speech-spoof-bench==0.3.4
+  match: scoring
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-06-10'
+notes: AASIST3 (KAN-enhanced AASIST + wav2vec 2.0 XLS-R-53 SSL front-end) from mtuciru/AASIST3, weights
+  from lab260/AASIST3 (model.safetensors, self-contained incl. the SSL encoder). Loaded via PyTorchModelHubMixin.from_pretrained;
+  the XLS-R architecture is built from the wav2vec2-large-xlsr-53 config then every weight is overwritten
+  by the checkpoint. Preemphasis (0.97) applied to the full waveform before a deterministic first-64600-sample
+  window, matching the source datasets/generic.py eval pipeline. score = logit for class 1 (bona fide);
+  higher = more bona fide (source label map bonafide=1, validation uses outputs[:, 1]).