issf/spoof-superb-scores

Spoof-SUPERB: Detection Score Files

Raw detection scores for every model and every evaluation corpus in Spoof-SUPERB, a benchmark of self-supervised speech representations for audio deepfake detection.

Every number in the paper can be recomputed from these files alone, with no GPU, no model checkpoints and no access to the underlying audio corpora.

• Code: https://github.com/issflab/spoof_SUPERB
• Leaderboard: https://huggingface.co/spaces/issf/Spoof-SUPERB

What this is

One score file per (model, corpus) pair. Each row is one trial:

utterance_id  source  key  score

key is bonafide or spoof; score is the model's raw detection score, where a higher value means more genuine. Equal error rate is computed directly from these two columns, so the published tables and figures follow from this data without rerunning any model.

Layout

raw/
  linear_head/          19 self-supervised front-ends, frozen, with a linear head
    asvspoof2019_la_eval/
    asvspoof2021_la/
    asvspoof2021_df/
    asvspoof5/
    asvspoof_ld/
    deepfake_eval_2024_segmented/
    famous_figures/
    in_the_wild/
    spoofceleb/
    mailabs/
    mlaad_v10/          per-system scores behind the synthesis-diversity analysis
  non_ssl/              LFCC-GMM and AASIST reference systems, same corpora
manifest.json           per-file sha256, row counts and score quantiles

277 files, 7.5 GB.

Verifying integrity

manifest.json carries a sha256, row count, bonafide/spoof counts and score quantiles for each indexed file. To check a download:

import json, hashlib
m = json.load(open("manifest.json"))
# each entry: {"path": ..., "sha256": ..., "n_rows": ..., "eer_percent": ...}

Note the manifest indexes 234 of the 277 files. The MLAAD per-system scores and part of the non-SSL baseline set are published here but not indexed by it; they are covered by the repository's own analysis code.

Reproducing the paper

git clone https://github.com/issflab/spoof_SUPERB
# point the repo's score root at your download of this dataset, then:
bash bin/reproduce_main_results.sh

What is NOT here

The audio corpora themselves. Spoof-SUPERB evaluates on ASVspoof 2019 LA, ASVspoof 2021 LA/DF, ASVspoof 5, the ASVspoof Laundered Database, DeepfakeEval 2024, In-the-Wild, Famous Figures, SpoofCeleb, and MLAAD with M-AILABS. Each is distributed by its own authors under its own licence and must be obtained from them. These score files contain only utterance identifiers and model outputs, never audio.

Licence

The score files are released under CC-BY-4.0. The underlying corpora remain under the licences of their respective authors, which this release does not alter.

Citation

@article{ali2026spoofsuperb,
  title  = {Spoof-SUPERB: A Comprehensive Benchmark of Self-Supervised Speech
            Representations for Audio Deepfake Detection},
  author = {Ali, Hashim and Adupa, Nithin Sai and Malik, Hafiz},
  year   = {2026}
}