Add whisper-mfcc-mesonet LRLspoof submission (srr_complement 47.62%)

submissions/whisper-mfcc-mesonet.yaml

@@ -0,0 +1,52 @@
+schema_version: 5
+system:
+  name: Whisper-MFCC-MesoNet
+  slug: whisper-mfcc-mesonet
+  description: '(Whisper + MFCC) MesoNet anti-spoofing countermeasure: a Whisper tiny.en
+    audio encoder front-end concatenated (2 channels) with an MFCC+Δ+ΔΔ front-end,
+    feeding a MesoInception4 classifier. Whisper encoder fine-tuned end-to-end. FP32.
+    Upstream eval pipeline reproduced: sox silence-trim (silence 1 0.2 1% -1 0.2 1%)
+    then a 30 s (480000-sample) repeat-pad window. This is the best MesoNet configuration
+    from the paper.'
+  code: https://github.com/piotrkawa/deepfake-whisper-features
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/WhisperMFCCMesoNet
+  params_millions: 7.660881
+  paper:
+    arxiv_id: '2306.01428'
+    url: https://arxiv.org/abs/2306.01428
+    bibtex: "@inproceedings{kawa23b_interspeech,\n  title     = {Improved DeepFake\
+      \ Detection Using Whisper Features},\n  author    = {Piotr Kawa and Marcin Plata\
+      \ and Micha{\\l} Czuba and Piotr Szyma{\\'n}ski and Piotr Syga},\n  year   \
+      \   = {2023},\n  booktitle = {Proc. INTERSPEECH 2023},\n  pages     = {4009--4013},\n\
+      \  doi       = {10.21437/Interspeech.2023-1537},\n}\n"
+dataset:
+  id: lab260/LRLspoof
+  revision: 5caf11349cf96aa198527cfb040cdba585cdd49e
+  split: test
+scores:
+  srr_complement: 47.617064966273546
+  n_trials: 1304455
+  n_skipped: 286
+calibration:
+  source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
+  threshold: 0.8410320616949153
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/WhisperMFCCMesoNet/resolve/332d5754dfcced925e4f61fba11727fc7c0f1f7e/.eval_results/lab260/LRLspoof/scores.txt
+  scores_sha256: 3ba62be13dfeb9e6cf4a43bed62777f450dda5abedd7504c4af27ccdaff50854
+  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: Whisper-MFCC-MesoNet 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 31.34%), transferred via the calibration block. Verified
+  with reproduce --scoring.