Add tropa-voice-mini submission for DECRO (EER=3.7325)

submissions/tropa-voice-mini.yaml

@@ -0,0 +1,52 @@
+schema_version: 4
+system:
+  name: tropa-voice-mini
+  slug: tropa-voice-mini
+  description: 'tropa-voice-mini (wasitaigenerated.com, 2026): the open-weights AI-voice detector behind
+    the wasitaigenerated API''s small tier. A wav2vec 2.0 XLS-R 300M encoder fine-tuned end-to-end, with
+    a multi-layer attention head (SLS head): softmax gate over all 25 hidden states, LayerNorm + Linear(1024,256)
+    + GELU, attentive statistics pooling, 2-layer MLP to one AI logit. Per-window zero-mean / unit-std
+    waveform normalisation inside the model. Authors'' released checkpoint; training data undisclosed.
+    Input: 16 kHz, first 64600 samples (Arena 4 s cap), shorter clips repeat-padded (the authors score
+    6 s windows with a 3 s hop over the whole clip and calibrate the top-30 % mean logit). score = -(pre-calibration
+    AI logit), higher = human. FP32 (TF32 off).'
+  code: https://huggingface.co/wasitaigeneratedcom/ai-voice-detector-small/blob/4858b4064b1f8530def904e6340d2780ad847160/detector.py
+  checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/tropa-voice-mini/blob/c57dd5ffc505ee75c8109469d0e11fa1239d3b6f/model.safetensors
+  params_millions: 315.867779
+dataset:
+  id: SpeechAntiSpoofingBenchmarks/DECRO
+  revision: e54c29df5b81bac75e5cc294d1ab0d4db72a5c71
+  split: test
+scores:
+  eer_percent: 3.732480761366593
+  n_trials: 37314
+  n_skipped: 0
+artifact:
+  scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/tropa-voice-mini/resolve/04ca32fe5a0cfffc41fbba225ed30e2088d1a286/.eval_results/SpeechAntiSpoofingBenchmarks/DECRO/scores.txt
+  scores_sha256: a962c12912594bf72fb5415844d94d897a7c0b9e44091768336409ebb3950bab
+  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: 'tropa-voice-mini = HF wasitaigeneratedcom/ai-voice-detector-small @ 4858b4064b1f8530def904e6340d2780ad847160,
+  model.safetensors (sha256 0fa2240e5178e2e06853ab2444b7188ce72e66c12adef87e3229bdd9be5ea2f1; 436 fp32
+  tensors, strict load, all used) + config.json (vendored as tropa_config.json; the XLS-R 300M architecture
+  is built from it, no Hub download). Apache-2.0 per the model card (no LICENSE file in the repo) -> weights
+  re-hosted. Model classes (AttentiveStatsPool, SLSHead, VoiceDetectorModel) vendored verbatim from the
+  repo''s detector.py (sha256 1ccce01f..., read, never imported from the Hub cache). Verified against
+  the authors'' VoiceDetector (its window logic + model, same 4 s inputs): max |d| 3.8e-6 on 66 J-SPAW_LA
+  clips, and their calibrated ai_score is reproduced exactly from our score (ai_score = sigmoid(1.4468
+  * logit + 1.3163), logit = -score). Deviations: Arena 4 s cap (first 64600 samples, repeat-pad) instead
+  of the authors'' 6 s windows / 3 s hop / native-length short clips with top-30 % aggregation; the monotone
+  calibration is dropped (rank-irrelevant); strict FP32 (TF32 off). Training data is undisclosed (the
+  card only says its own Commercial-Voices-2026-mini speakers/texts are unseen in training) -> no trained_on
+  entries, but leakage cannot be ruled out; the card itself benchmarks In-the-Wild and SONAR. No paper
+  -> unpublished tier. Self-reported EER% (authors'' sliding-window protocol): In-the-Wild 0.98 / SONAR
+  0.18 / Deepfake-Eval-2024 13.5 / MLAAD-tiny 2.33 / Commercial-Voices-2026-mini 0.19. J-SPAW_LA EER 0.125%
+  (PyTorch; TensorRT FP32 identical on RTX 4070 Ti SUPER and RTX 5060 Ti, Spearman 0.999995).'