Add tropa-voice-mini submission for LRLspoof (1-SRR=2.9281)

#42
Files changed (1) hide show
  1. submissions/tropa-voice-mini.yaml +55 -0
submissions/tropa-voice-mini.yaml ADDED
@@ -0,0 +1,55 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 4
2
+ system:
3
+ name: tropa-voice-mini
4
+ slug: tropa-voice-mini
5
+ description: 'tropa-voice-mini (wasitaigenerated.com, 2026): the open-weights AI-voice detector behind
6
+ the wasitaigenerated API''s small tier. A wav2vec 2.0 XLS-R 300M encoder fine-tuned end-to-end, with
7
+ a multi-layer attention head (SLS head): softmax gate over all 25 hidden states, LayerNorm + Linear(1024,256)
8
+ + GELU, attentive statistics pooling, 2-layer MLP to one AI logit. Per-window zero-mean / unit-std
9
+ waveform normalisation inside the model. Authors'' released checkpoint; training data undisclosed.
10
+ Input: 16 kHz, first 64600 samples (Arena 4 s cap), shorter clips repeat-padded (the authors score
11
+ 6 s windows with a 3 s hop over the whole clip and calibrate the top-30 % mean logit). score = -(pre-calibration
12
+ AI logit), higher = human. FP32 (TF32 off).'
13
+ code: https://huggingface.co/wasitaigeneratedcom/ai-voice-detector-small/blob/4858b4064b1f8530def904e6340d2780ad847160/detector.py
14
+ checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/tropa-voice-mini/blob/c57dd5ffc505ee75c8109469d0e11fa1239d3b6f/model.safetensors
15
+ params_millions: 315.867779
16
+ dataset:
17
+ id: lab260/LRLspoof
18
+ revision: ac76f0e5472fdb68a6726e1fa94a148316e76916
19
+ split: test
20
+ scores:
21
+ srr_complement: 2.9280714385942264
22
+ n_trials: 1304455
23
+ n_skipped: 286
24
+ calibration:
25
+ source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
26
+ threshold: 1.8255537514124294
27
+ artifact:
28
+ scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/tropa-voice-mini/resolve/09463a0b721f6c490591a24eba551520a45b12c7/.eval_results/lab260/LRLspoof/scores.txt
29
+ scores_sha256: 887864aceefbce9fb34329f703f5ebfda9e99a37e6a6028963fe0e4defefcc3d
30
+ bench_version: speech-spoof-bench==0.4.1
31
+ reproduction:
32
+ reproduced_by: SpeechAntiSpoofingBenchmarks
33
+ reproduced_at: '2026-11-16'
34
+ reproduced_bench_version: speech-spoof-bench==0.4.1
35
+ match: inference
36
+ submitter:
37
+ hf_username: korallll
38
+ contact: k.n.borodin@mtuci.ru
39
+ submitted_at: '2026-11-16'
40
+ notes: 'tropa-voice-mini = HF wasitaigeneratedcom/ai-voice-detector-small @ 4858b4064b1f8530def904e6340d2780ad847160,
41
+ model.safetensors (sha256 0fa2240e5178e2e06853ab2444b7188ce72e66c12adef87e3229bdd9be5ea2f1; 436 fp32
42
+ tensors, strict load, all used) + config.json (vendored as tropa_config.json; the XLS-R 300M architecture
43
+ is built from it, no Hub download). Apache-2.0 per the model card (no LICENSE file in the repo) -> weights
44
+ re-hosted. Model classes (AttentiveStatsPool, SLSHead, VoiceDetectorModel) vendored verbatim from the
45
+ repo''s detector.py (sha256 1ccce01f..., read, never imported from the Hub cache). Verified against
46
+ the authors'' VoiceDetector (its window logic + model, same 4 s inputs): max |d| 3.8e-6 on 66 J-SPAW_LA
47
+ clips, and their calibrated ai_score is reproduced exactly from our score (ai_score = sigmoid(1.4468
48
+ * logit + 1.3163), logit = -score). Deviations: Arena 4 s cap (first 64600 samples, repeat-pad) instead
49
+ of the authors'' 6 s windows / 3 s hop / native-length short clips with top-30 % aggregation; the monotone
50
+ calibration is dropped (rank-irrelevant); strict FP32 (TF32 off). Training data is undisclosed (the
51
+ card only says its own Commercial-Voices-2026-mini speakers/texts are unseen in training) -> no trained_on
52
+ entries, but leakage cannot be ruled out; the card itself benchmarks In-the-Wild and SONAR. No paper
53
+ -> unpublished tier. Self-reported EER% (authors'' sliding-window protocol): In-the-Wild 0.98 / SONAR
54
+ 0.18 / Deepfake-Eval-2024 13.5 / MLAAD-tiny 2.33 / Commercial-Voices-2026-mini 0.19. J-SPAW_LA EER 0.125%
55
+ (PyTorch; TensorRT FP32 identical on RTX 4070 Ti SUPER and RTX 5060 Ti, Spearman 0.999995).'