Add aasist-l LRLspoof submission (srr_complement 31.86%)

submissions/aasist-l.yaml

@@ -0,0 +1,54 @@
+schema_version: 5
+system:
+  name: AASIST-L
+  slug: aasist-l
+  description: 'AASIST-L: the lightweight variant of AASIST (audio anti-spoofing using
+    integrated spectro-temporal graph attention networks). Same architecture as AASIST
+    — sinc-convolution front-end, RawNet2-style residual encoder, and heterogeneous
+    stacking graph attention over spectral and temporal sub-graphs with a learnable
+    readout — but with a narrower residual stack and graph dimensions (~85k params
+    vs AASIST''s ~298k). Official clovaai/aasist ASVspoof2019 LA pretrained checkpoint,
+    FP32, deterministic first-64600-sample window (no random crop).'
+  code: https://github.com/clovaai/aasist
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/AASIST-L/blob/e4185b270ec20077c918e06a45093717a1bd5e30/AASIST-L.pth
+  params_millions: 0.085306
+  paper:
+    arxiv_id: '2110.01200'
+    url: https://arxiv.org/abs/2110.01200
+    bibtex: "@inproceedings{jung2022aasist,\n  title={{AASIST}: Audio Anti-Spoofing\
+      \ Using Integrated Spectro-Temporal Graph Attention Networks},\n  author={Jung,\
+      \ Jee-weon and Heo, Hee-Soo and Tak, Hemlata and Shim, Hye-jin and Chung, Joon\
+      \ Son and Lee, Bong-Jin and Yu, Ha-Jin and Evans, Nicholas},\n  booktitle={ICASSP\
+      \ 2022 - 2022 IEEE International Conference on Acoustics, Speech and Signal\
+      \ Processing (ICASSP)},\n  pages={6367--6371},\n  year={2022},\n  organization={IEEE}\n\
+      }\n"
+dataset:
+  id: lab260/LRLspoof
+  revision: 5caf11349cf96aa198527cfb040cdba585cdd49e
+  split: test
+scores:
+  srr_complement: 31.86021136831193
+  n_trials: 1304455
+  n_skipped: 286
+calibration:
+  source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
+  threshold: -0.41200456050955414
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/AASIST-L/resolve/9b581e4690b582a71f44140cc1d407defe6d1a4a/.eval_results/lab260/LRLspoof/scores.txt
+  scores_sha256: 8e26da4e5058013d7c84f433b9368316f12994f5e5d9b4f8961c78ffe4a93ea3
+  bench_version: speech-spoof-bench==0.4.1
+reproduction:
+  reproduced_by: SpeechAntiSpoofingBenchmarks
+  reproduced_at: '2026-06-22'
+  reproduced_bench_version: speech-spoof-bench==0.4.1
+  match: scoring
+submitter:
+  hf_username: korallll
+  contact: k.n.borodin@mtuci.ru
+submitted_at: '2026-06-22'
+notes: AASIST-L on LRLspoof (spoof-only, srr_complement = 1-SRR, lower better). TensorRT
+  FP32 engine (parity-verified vs PyTorch). Scored 1,304,455 of 1,304,455 TTS utterances;
+  286 source files are empty/zero-length (skipped), 172 corrupt-header WAVs recovered
+  via ffmpeg. Audio resampled per-file to 16 kHz; score = net(x)[:,1] (higher = bona
+  fide). t* = this model's own DeepVoice EER operating-point threshold (DeepVoice
+  EER 49.68%), transferred via the calibration block. Verified with reproduce --scoring.