Datasets:
Add Spectra-AASIST3 submission (CFAD) (#4)
Browse files- Add Spectra-AASIST3 submission (CFAD) (089ecc74b63baf19937c2c9eca7f6dea70a5acb2)
- Stamp reproduction for Spectra-AASIST3 CFAD (ab27d80f43beeab241af85b5a3f7ca4e673c844c)
submissions/spectra-aasist3.yaml
ADDED
|
@@ -0,0 +1,38 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
schema_version: 4
|
| 2 |
+
system:
|
| 3 |
+
name: Spectra-AASIST3
|
| 4 |
+
slug: spectra-aasist3
|
| 5 |
+
description: 'Speech deepfake detector: a wav2vec 2.0 XLS-R-300m self-supervised front-end feeding a
|
| 6 |
+
single-layer MLP bridge (1024->128) into a KAN-enhanced AASIST (KAN-AASIST) 2-class classifier with
|
| 7 |
+
four-branch spectro-temporal graph attention and a KAN output layer. Pre-release lab260/Spectra-AASIST3
|
| 8 |
+
checkpoint. FP32, preemphasis (0.97), deterministic first-64600-sample window (no random crop). score
|
| 9 |
+
= output logit for class 1 (bona fide). Unpublished / pre-release model (no paper).
|
| 10 |
+
|
| 11 |
+
'
|
| 12 |
+
code: https://huggingface.co/lab260/Spectra-AASIST3
|
| 13 |
+
checkpoint: https://huggingface.co/lab260/Spectra-AASIST3
|
| 14 |
+
params_millions: 318.9489
|
| 15 |
+
dataset:
|
| 16 |
+
id: SpeechAntiSpoofingBenchmarks/CFAD
|
| 17 |
+
revision: 53d7855c1c378524f7b7b1030bcb6b2caa327fe6
|
| 18 |
+
split: test
|
| 19 |
+
scores:
|
| 20 |
+
eer_percent: 0.7095238095238096
|
| 21 |
+
n_trials: 62999
|
| 22 |
+
n_skipped: 0
|
| 23 |
+
artifact:
|
| 24 |
+
scores_url: https://huggingface.co/lab260/Spectra-AASIST3/resolve/1e56e594fe298ab0d886247f17bda1224fc5ffa7/.eval_results/SpeechAntiSpoofingBenchmarks/CFAD/scores.txt
|
| 25 |
+
scores_sha256: 605895e8f539aa9ee70f9ccd45a370be2ee586f476f61f03a0aeebb3e8577ca0
|
| 26 |
+
bench_version: speech-spoof-bench==0.3.4
|
| 27 |
+
submitter:
|
| 28 |
+
hf_username: korallll
|
| 29 |
+
contact: k.n.borodin@mtuci.ru
|
| 30 |
+
submitted_at: '2026-06-06'
|
| 31 |
+
notes: Spectra-AASIST3 (wav2vec2 XLS-R-300m SSL front-end + MLP bridge + KAN-AASIST). Preemphasis 0.97,
|
| 32 |
+
deterministic first-64600-sample window. score = logit class 1 (bona fide). Unpublished/pre-release
|
| 33 |
+
(no paper). Scores hosted in lab260/Spectra-AASIST3/.eval_results.
|
| 34 |
+
reproduction:
|
| 35 |
+
reproduced_by: SpeechAntiSpoofingBenchmarks
|
| 36 |
+
reproduced_at: '2026-06-07'
|
| 37 |
+
reproduced_bench_version: speech-spoof-bench==0.3.4
|
| 38 |
+
match: scoring
|