Benchmark-ready packaging of the Round 1 (R1) evaluation partition of Track 1.2 (Fake Game Detection / FG-D) from the ADD 2023 challenge — the Second Audio Deepfake Detection Challenge (arXiv 2305.13774). Binary anti-spoofing over Mandarin audio: bonafide (genuine human speech) vs. spoof (synthesized / fake speech). Track 1.2 is the defending side of an attack-and-defense game; the fakes are produced by the attacking party (Track 1.1). The bonafide speech derives from the AISHELL-3 multi-speaker Mandarin corpus.

⚠️ Labels only — the audio is not redistributed here

The ADD 2023 audio is licensed CC BY-NC-ND 4.0 (NonCommercial-NoDerivatives), which does not permit us to redistribute the waveforms. This repo therefore ships only data/labels.parquet (utterance_id + label) — no audio at all. That is everything the Arena needs: scoring is reproduced from a model's scores.txt plus these labels and never transfers audio.

load_dataset(...) returns only the id/label table here (the viewer is disabled); it is not the audio access path.

How to obtain the audio

Download the original ADD 2023 Track 1.2 audio from the source release:

Zenodo: https://zenodo.org/records/12151404

Extract the Round 1 test split into a directory of ADD2023_T1.2R1_E_*.wav files (16 kHz mono WAV). The accompanying protocol is label.txt (<file>.wav <genuine|fake>), which is exactly what data/labels.parquet was derived from (see _build_labels.py).

How to compute scores locally

Once you have licensed access to the audio, run your anti-spoofing model over the local audio directory and emit a scores.txt (<utterance_id> <score>, higher = more bonafide), then submit it to the Arena (the labels here verify it) — see the package's docs/submitting/. _build_labels.py documents the exact id/label derivation.

Schema (`data/labels.parquet`)

Column	Type	Description
utterance_id	string	Audio filename stem, e.g. `ADD2023_T1.2R1_E_00000000`
label	int8	`0` = bonafide (genuine), `1` = spoof (fake)

utterance_id is the audio file's stem (no .wav). A submitter's scores.txt keys by this id.

Stats

Stat	Value
Total trials	111976
Bonafide (genuine)	80000
Spoof (fake)	31976

Arena scoring

Standard EER (eer_percent, lower is better), computed over all 111 976 utterances. The seeded random-baseline scores ≈ 50 % EER by construction.

Source & citation

Original audio: https://zenodo.org/records/12151404 (CC BY-NC-ND 4.0)
Protocol: label.txt (<file>.wav <genuine|fake>)
Paper: ADD 2023 — the Second Audio Deepfake Detection Challenge, arXiv 2305.13774

@inproceedings{yi2023add,
  title     = {{ADD} 2023: the Second Audio Deepfake Detection Challenge},
  author    = {Yi, Jiangyan and Tao, Jianhua and Fu, Ruibo and Yan, Xinrui and
               Wang, Chenglong and Wang, Tao and Zhang, Chu Yuan and Zhang, Xiaohui and
               Zhao, Yan and Ren, Yong and Xu, Le and Zhou, Junzuo and Gu, Hao and
               Wen, Zhengqi and Liang, Shan and Lv, Zheng and Xu, Bin and others},
  booktitle = {Proc. DADA@IJCAI},
  pages     = {125--130},
  year      = {2023}
}