Add SSL-Mamba submission for LRLspoof (1-SRR=23.4563)

submissions/ssl-mamba.yaml

@@ -0,0 +1,56 @@
+schema_version: 4
+system:
+  name: SSL-Mamba
+  slug: ssl-mamba
+  description: 'SSL-Mamba (Tran et al., Interspeech 2025): a fully fine-tuned wav2vec 2.0 XLS-R 300M front-end
+    whose 25 hidden states are each projected to 128-d (Linear + SiLU) and summed (self-gated layer aggregation),
+    a multi-kernel gated convolution (MultiConv cgMLP, kernels 7/15/23/31) fusion module, 12 unidirectional
+    Mamba (v1 selective SSM) residual blocks with LayerNorm, multi-head attention pooling (4 heads, mean
+    + std) and a 2-way head. Official authors'' checkpoint (interspeech2025.pt), trained on ASVspoof2019
+    LA train. Not to be confused with XLSR-Mamba (Xiao & Das), a different model.'
+  code: https://github.com/hoanmyTran/spoofing_ssl_mamba
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/SSL-Mamba/blob/3565084c8f0125b48ca79525674a05167638e40a/model.safetensors
+  params_millions: 317.5213
+  paper:
+    arxiv_id: 10.21437/Interspeech.2025-1703
+    url: https://www.isca-archive.org/interspeech_2025/tran25b_interspeech.html
+    bibtex: "@inproceedings{tran25b_interspeech,\n  title     = {{Leveraging SSL Speech Features and Mamba\
+      \ for Enhanced DeepFake Detection}},\n  author    = {Tran, Hoan My and Lolive, Damien and Guennec,\
+      \ David and Sini, Aghilas and Delhay, Arnaud and Marteau, Pierre-Fran{\\c{c}}ois},\n  booktitle\
+      \ = {Interspeech 2025},\n  pages     = {5323--5327},\n  year      = {2025},\n  doi       = {10.21437/Interspeech.2025-1703}\n\
+      }\n"
+dataset:
+  id: lab260/LRLspoof
+  revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
+  split: test
+scores:
+  srr_complement: 23.45631586090453
+  n_trials: 1304455
+  n_skipped: 286
+calibration:
+  source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
+  threshold: -1.905060580338983
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/SSL-Mamba/resolve/27cce04fa34385a399e45c1a7d225bf35f30ed1d/.eval_results/lab260/LRLspoof/scores.txt
+  scores_sha256: 38d692d26537ef5799b9bd3bc67b54192f113a76bf06590d494ab9f42fdf15b0
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-11-04'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: inference
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-11-04'
+notes: 'SSL-Mamba (Tran et al., Interspeech 2025, doi 10.21437/Interspeech.2025-1703); code github hoanmyTran/spoofing_ssl_mamba
+  @ 9afc33de7f74a9b9b3423a9028c10e1d18c1c6b8 (MIT); weights HF hoanmyTran/ssl_spoofing_mamba @ 4e72dfcef4f03460e08f68a4a5fbcf3705a3414e
+  interspeech2025.pt (Apache-2.0), converted to safetensors (training-only loss_fn.weight dropped), strict
+  load, 317,521,282 parameters. Trained on ASVspoof2019 LA train. Deterministic first-64600-sample window
+  (4.04 s, the source training max_len), shorter clips repeat-tiled (source dataset.pad); no resampling
+  or normalisation; FP32; score = log_softmax[:,1] - log_softmax[:,0] (bona fide = class 1, source llr).
+  Deviations: the authors evaluate on the full variable-length utterance -- here a fixed 4 s window is
+  used for every set (maintainer decision; the paper reports both, e.g. 21LA fixed 2.77 / 21DF fixed 1.97
+  % EER); Mamba selective scan and causal depthwise conv in plain PyTorch (same math as the mamba_ssm
+  CUDA kernels; chunked scan == selective_scan_ref within 1e-4); XLS-R built from its config (no Hub weights,
+  the checkpoint holds the fine-tuned front-end); espnet LayerNorm/activations vendored.'