| schema_version: 4 |
|
|
| system: |
| name: random-baseline |
| slug: random-baseline |
| description: > |
| Reference random baseline. Returns N(0, 1) for every utterance using a |
| fixed seed (seed=0). EER ≈ 50% by construction. Used as the seeded |
| smoke-test baseline for the arena (roadmap Phase 3). |
| code: https://github.com/SpeechAntiSpoofingBenchmarks/speech-spoof-bench |
| checkpoint: https://huggingface.co/SpeechAntiSpoofingBenchmarks/random-baseline-asas |
| params_millions: 1 |
| paper: |
| arxiv_id: "1911.01601" |
| url: https://arxiv.org/abs/1911.01601 |
| bibtex: | |
| @article{wang2020asvspoof, |
| title={ASVspoof 2019: A large-scale public database of synthesized, |
| converted and replayed speech}, |
| author={Wang, Xin and Yamagishi, Junichi and Todisco, Massimiliano and |
| Delgado, H{\'e}ctor and Nautsch, Andreas and Evans, Nicholas |
| and Sahidullah, Md and Vestman, Ville and Kinnunen, Tomi and |
| Lee, Kong Aik and others}, |
| journal={Computer Speech \& Language}, |
| volume={64}, |
| pages={101114}, |
| year={2020}, |
| publisher={Elsevier} |
| } |
| |
| dataset: |
| id: SpeechAntiSpoofingBenchmarks/PyAra |
| revision: 40167badd6d8a289f02054761ee5411f49227b87 |
| split: test |
|
|
| scores: |
| eer_percent: 49.85934251117785 |
| n_trials: 201778 |
| n_skipped: 0 |
|
|
| artifact: |
| scores_url: https://huggingface.co/SpeechAntiSpoofingBenchmarks/random-baseline-asas/resolve/2a2743fcf3f916ed080232f19196706accd1a543/.eval_results/SpeechAntiSpoofingBenchmarks/PyAra/scores.txt |
| scores_sha256: d237bbf69110e7938f586d6653aaa4e45b2369a4f763aac5ed69e8cd7f118b57 |
| bench_version: speech-spoof-bench==0.3.4 |
|
|
| reproduction: |
| reproduced_by: SpeechAntiSpoofingBenchmarks |
| reproduced_at: 2026-06-10 |
| reproduced_bench_version: speech-spoof-bench==0.3.4 |
| match: scoring |
|
|
| submitter: |
| hf_username: SpeechAntiSpoofingBenchmarks |
| contact: k.n.borodin@mtuci.ru |
|
|
| submitted_at: 2026-06-10 |
|
|
| notes: > |
| Seeded random baseline for the newly-added PyAra dataset (Russian audio |
| deepfake detection, https://www.kaggle.com/datasets/alep079/pyara; |
| CC BY-NC-SA 4.0). 201,778 trials: 73,583 bonafide / 128,195 spoof from 5 |
| generation algorithms (alg_1..alg_5). Random baseline has no paper of its own; |
| the paper field cites the ASVspoof 2019 dataset paper as a placeholder, per the |
| Phase 3 convention and consistent with the other random-baseline submissions |
| (same system slug). Reproduction verified by the maintainer via |
| `reproduce --scoring` (scores_sha256 matched; Δ EER 0.0; 201778 trials / 0 |
| skipped) at speech-spoof-bench==0.3.4. |
| |
