README.md

---
license: cc-by-4.0
language:
  - zh
pretty_name: CFAD
task_categories: [audio-classification]
size_categories: [10K<n<100K]
configs:
  - config_name: default
    data_files:
      - {split: test, path: "data/test-*.parquet"}
tags:
  - anti-spoofing
  - audio-deepfake-detection
  - speech
  - benchmark
  - arena-ready
arxiv:
  - "2207.12308"
---

# CFAD

Benchmark-ready packaging of the **CFAD** (Chinese Fake Audio Detection) **clean test**
set ([arXiv 2207.12308](https://arxiv.org/abs/2207.12308)), for speech anti-spoofing and
synthetic / deepfake voice detection on Mandarin Chinese speech.

## Overview

CFAD is a large-scale Chinese fake-audio detection corpus. This repo packages the **clean**
version's two test partitions:

- **`test_seen`** — spoof systems and real corpora **also present** in the train/dev splits.
- **`test_unseen`** — spoof systems and real corpora held **out** of training, to measure
  generalisation to unseen attacks.

The label is **binary**: **bonafide** (genuine human speech) vs. **spoof** (synthesized /
vocoded speech). It is derived from the source directory layout — clips under `fake_clean/`
are spoof; clips under `real_clean/` are bonafide. There is no separate protocol file.

| Partition | Side | Sources | n |
|-----------|------|---------|---|
| test_seen | bonafide | aishell1, aishell3, magicread, thchs30 | 14,000 |
| test_seen | spoof | gl, hifigan, lpcnet, mbmelgan, pwg, straight, stylegan, wavenet | 28,000 |
| test_unseen | bonafide | magicconversa, selfrecording | 7,000 |
| test_unseen | spoof | fasthifigan, partiallyfake, tacohifigan, world | 14,000 |

One aishell1 outlier (`BAC009S0764W0123.wav`, a 178 MB / ~97 min concatenation artifact) is
dropped, so the bonafide count is **20,999** rather than 21,000.

## License & redistribution

CFAD is released under the **Creative Commons Attribution 4.0 International (CC BY 4.0)**
license, which permits redistribution with attribution — see `LICENSE.txt`. Audio is
re-encoded to 16 kHz mono FLAC for a uniform schema; labels are unmodified.

## Schema

| Column | Type | Description |
|--------|------|-------------|
| `path` | `string` | source-relative path (e.g. `test_seen_clean/fake_clean/gl/SSB07800001_gl.wav`), unique |
| `audio` | `Audio(16000)` | 16 kHz mono FLAC (re-encoded from the 16 kHz PCM source wav) |
| `label` | `ClassLabel` | `"bonafide"` (0) / `"spoof"` (1) |
| `notes` | `string` | JSON: `utterance_id`, `partition`, `source` |

`notes` example:
```json
{"utterance_id": "CFAD_test_seen_clean_fake_clean_gl_SSB07800001_gl", "partition": "test_seen", "source": "gl"}
```

## Quick Start

```python
from datasets import load_dataset

ds = load_dataset("SpeechAntiSpoofingBenchmarks/CFAD", split="test")
print(ds[0])
```

## Stats

| Stat | Value |
|------|-------|
| Total trials | 62,999 |
| Bonafide | 20,999 |
| Spoof | 42,000 |

## Source provenance

- Paper: https://arxiv.org/abs/2207.12308
- Scope: the **clean** version's `test_seen_clean` + `test_unseen_clean` partitions
  (the `codec`/`noisy` robustness variants and train/dev splits are not included).
- Labels derived from the source directory layout (`fake_clean/` = spoof, `real_clean/` =
  bonafide).

## Evaluation

For evaluation instructions and submission format, see [`submissions/README.md`](submissions/README.md).

## Citation

> **Paper:** Haoxin Ma et al., *CFAD: A Chinese Dataset for Fake Audio Detection*,
> [arXiv:2207.12308](https://arxiv.org/abs/2207.12308) (2022).

```bibtex
@article{ma2022cfad,
  title   = {CFAD: A Chinese Dataset for Fake Audio Detection},
  author  = {Ma, Haoxin and Yi, Jiangyan and Wang, Chenglong and Yan, Xinrui and
             Tao, Jianhua and Wang, Tao and Wang, Shiming and Fu, Ruibo},
  journal = {arXiv preprint arXiv:2207.12308},
  year    = {2022}
}
```

## Maintainer

Contact: k.n.borodin@mtuci.ru