korallll commited on
Commit
f646d2e
·
verified ·
1 Parent(s): 1cbf1a7

Add spectra-0 LRLspoof submission (srr_complement 0.75%)

Browse files
Files changed (1) hide show
  1. submissions/spectra-0.yaml +44 -0
submissions/spectra-0.yaml ADDED
@@ -0,0 +1,44 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ schema_version: 5
2
+ system:
3
+ name: Spectra-0
4
+ slug: spectra-0
5
+ description: 'Speech deepfake detector: a wav2vec 2.0 XLS-R-300m self-supervised
6
+ front-end feeding a single-layer MLP bridge (1024->128, SELU) into an ECAPA-TDNN
7
+ 2-class classifier. Pre-release lab260/spectra_0 checkpoint. FP32, preemphasis
8
+ (0.97), deterministic first-64600-sample window (no random crop). score = output
9
+ logit for class 1 (bona fide). Unpublished / pre-release model (no paper).
10
+
11
+ '
12
+ code: https://huggingface.co/lab260/spectra_0
13
+ checkpoint: https://huggingface.co/lab260/spectra_0
14
+ params_millions: 318.3322
15
+ dataset:
16
+ id: lab260/LRLspoof
17
+ revision: 5caf11349cf96aa198527cfb040cdba585cdd49e
18
+ split: test
19
+ scores:
20
+ srr_complement: 0.7544285006663265
21
+ n_trials: 1304455
22
+ n_skipped: 289
23
+ calibration:
24
+ source_dataset: SpeechAntiSpoofingBenchmarks/DeepVoice
25
+ threshold: 0.9788363096045197
26
+ artifact:
27
+ scores_url: https://huggingface.co/lab260/spectra_0/resolve/98446d860311b2444c49ff457b91c9c4b87e90b7/.eval_results/lab260/LRLspoof/scores.txt
28
+ scores_sha256: 748dff3c081c7f7e418b4dfd23637a2c37f080c0bed6df677d484a35eae857aa
29
+ bench_version: speech-spoof-bench==0.4.1
30
+ reproduction:
31
+ reproduced_by: SpeechAntiSpoofingBenchmarks
32
+ reproduced_at: '2026-06-23'
33
+ reproduced_bench_version: speech-spoof-bench==0.4.1
34
+ match: scoring
35
+ submitter:
36
+ hf_username: korallll
37
+ contact: k.n.borodin@mtuci.ru
38
+ submitted_at: '2026-06-23'
39
+ notes: Spectra-0 on LRLspoof (spoof-only, srr_complement = 1-SRR, lower better). TensorRT
40
+ engine (parity-verified vs PyTorch). Scored 1,304,455 of 1,304,455 TTS utterances;
41
+ 289 source files are empty/zero-length (skipped), 172 corrupt-header WAVs recovered
42
+ via ffmpeg. Audio resampled per-file to 16 kHz; score = net(x)[:,1] (higher = bona
43
+ fide). t* = this model's own DeepVoice EER operating-point threshold (DeepVoice
44
+ EER 3.07%), transferred via the calibration block. Verified with reproduce --scoring.