Datasets:
File size: 4,182 Bytes
33561dd 010e2d2 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 | ---
license: cc-by-nc-nd-4.0
language:
- zh
pretty_name: ADD 2022 Track 3 R1 Test (labels only)
task_categories:
- audio-classification
size_categories:
- 100K<n<1M
viewer: false
configs:
- config_name: default
data_files:
- split: test
path: data/labels.parquet
tags:
- anti-spoofing
- audio-deepfake-detection
- speech
- benchmark
- arena-ready
arxiv:
- "2202.08433"
---
# ADD 2022 — Track 3 (Audio Fake Game), Round 1 Test · labels only
Benchmark-ready packaging of the **Round 1 (R1) evaluation** partition of **Track 3
(Audio Fake Game / FG)** from the **ADD 2022** challenge
([arXiv 2202.08433](https://arxiv.org/abs/2202.08433)). Binary anti-spoofing:
**bonafide** (genuine human speech) vs. **spoof** (synthesized / fake speech).
## ⚠️ Labels only — the audio is not redistributed here
The ADD 2022 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 2022 Track 3 audio from the source release:
- **Zenodo:** https://zenodo.org/records/12188035
Extract the R1 test split into a `track3test/` directory of `ADD_E3_*.wav` files (16 kHz
mono WAV). The accompanying protocol is `track3_R1_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):
```bash
python _score_add22.py \
--audio-dir /path/to/add22track31test/track3test \
--model random-baseline \
--out scores.txt
```
`_score_add22.py` is a small, model-pluggable driver: it lists the audio directory,
decodes each clip with `soundfile` (sorted by id — spinning-disk friendly), calls your
model's `score(audio, sr)`, and writes `<utterance_id> <score>`. Swap `--model` for your
own `module:Class` implementing the package's `SimpleAntiSpoofingModel` interface. Then
submit `scores.txt` to the Arena (the labels here verify it) — see the package's
`docs/submitting/`.
## Schema (`data/labels.parquet`)
| Column | Type | Description |
|--------|------|-------------|
| utterance_id | string | Audio filename **stem**, e.g. `ADD_E3_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 | 112861 |
| Bonafide (genuine) | 20776 |
| Spoof (fake) | 92085 |
## Arena scoring
Standard **EER** (`eer_percent`, lower is better), computed over all 112 861 utterances.
The seeded `random-baseline` scores ≈ 50 % EER by construction.
## Source & citation
- **Original audio:** https://zenodo.org/records/12188035 (CC BY-NC-ND 4.0)
- **Protocol:** `track3_R1_label.txt`
- **Paper:** ADD 2022 — the first Audio Deep Synthesis Detection Challenge,
[arXiv 2202.08433](https://arxiv.org/abs/2202.08433)
```bibtex
@inproceedings{yi2022add,
title = {{ADD} 2022: the first Audio Deep Synthesis Detection Challenge},
author = {Yi, Jiangyan and Fu, Ruibo and Tao, Jianhua and Nie, Shuai and
Ma, Haoxin and Wang, Chenglong and Wang, Tao and Tian, Zhengkun and
Bai, Ye and Fan, Cunhang and Liang, Shan and Wang, Shiming and
Zhang, Shuai and Yan, Xinrui and Xu, Le and Wen, Zhengqi and Li, Haizhou},
booktitle = {ICASSP},
pages = {9216--9220},
year = {2022}
}
```
## Maintainer
Maintained by Kirill Borodin (SpeechAntiSpoofingBenchmarks).
- Email: kborodin.research@gmail.com
- Telegram: [@korallll_ai](https://t.me/korallll_ai)
|