SPEECHFAKE

# SpeechFake

Copyright © 2025 Ant Group

## 1. Dataset Structure

```
SpeechFake/
|- BD/
| |- BigVGAN # Bilingual Dataset
| | |- xxx/xxx.wav
| | |- ...
| |- ...
|- MD/ # Multilingual Dataset
| |- CosyVoice
| | |- xxx/xxx.wav
| | |- ...
| |- ...
|- Real/ # Real Dataset
| |- Aishell1
| | |- xxx/xxx.wav
| | |- ...
| |- ...
|
|- metadata/
| |- BD/
| | |- TTS_xxx.csv # Metadata of each generator
| | |- ...
| |
| |- MD/
| | |- TTS_xxx_en.csv # Metadata of each generator for each language
| | |- ...
| |
| |- Real/
| | |- Aishell1.csv # Metadata of each real datasets
| | |- ...
| |
| |- experiments/ # Metadata of train/dev/test data used in experiments
| |- baseline/
| | |- train_all.csv
| | |- dev_all.csv
| | |- test_all.csv
| | |- ...
| |
| |- cross_generator/
| | |- train_tts.csv
| | |- dev_tts.csv
| | |- test_tts.csv
| | |- ...
| |
| |- cross_lingual/
| | |- train.csv
| | |- dev.csv
| | |- test_en.csv
| | |- ...
| |
| |- cross_speaker/
| |- train.csv
| |- test_same_spk.csv
| |- test_diff_spk.csv
| |- ...
|
|- LICENSE.txt
|- README.md
```

## 2. Audio Description

All audio files are stored in WAV format at 16 kHz sampling rate.

- `BD/`: Contains bilingual speech deepfakes and real audio in English and Chinese.
- `MD/`: Contains multilingual speech deepfakes in 46 languages.
- `Real/`: Contains real speech data sourced from LibriTTS, VCTK, AISHELL1, AISHELL3, and CommonVoice.

For detailed information and file lists, refer to the `metadata/` directory.

## 3. Metadata Format

All metadata files are provided in CSV format with the following columns:

1) `file`: Relative file path to the audio
2) `label`: `bonafide` or `spoof`
3) `generator`: `TTS`, `VC`, or `NV`
4) `model`: Name of the speech generation model
5) `speaker`: Speaker identity
6) `language`: Language code (e.g., `en`, `zh`, `es`, etc.)

For the data structure of metadata:

- `metadata/BD` and `metadata/MD` include metadata for each speech generator.
- `metadata/Real` includes metadata for each real datasets.
- `metadata/experiments` contains metadata for training, development, and testing splits used in various experiments described in the paper, including:
- baseline
- cross-generator
- cross-lingual
- cross-speaker

## 4. License

The SpeechFake dataset is released under the CC-BY-4.0 License.

Please read `LICENSE.txt` for full details.

## 5. Citation

If you use this dataset in your work, please cite the following paper:

```
@inproceedings{huang2025speechfake,
title={SpeechFake: A Large-Scale Multilingual Speech Deepfake Dataset Incorporating Cutting-Edge Generation Methods},
author={Huang, Wen and Gu, Yanmei and Wang, Zhiming and Zhu, Huijia and Qian, Yanmin},
booktitle={Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
pages={9985--9998},
year={2025}
}
```

# 语音伪造（SpeechFake）

版权所有 © 2025 蚂蚁集团（Ant Group）

## 1. 数据集结构

SpeechFake/
|- BD/
| |- BigVGAN # 双语数据集
| | |- xxx/xxx.wav
| | |- ...
| |- ...
|- MD/ # 多语言数据集
| |- CosyVoice
| | |- xxx/xxx.wav
| | |- ...
| |- ...
|- Real/ # 真实数据集
| |- Aishell1
| | |- xxx/xxx.wav
| | |- ...
| |- ...
|
|- metadata/
| |- BD/
| | |- TTS_xxx.csv # 各生成器的元数据
| | |- ...
| |
| |- MD/
| | |- TTS_xxx_en.csv # 各语言对应生成器的元数据
| | |- ...
| |
| |- Real/
| | |- Aishell1.csv # 各真实数据集的元数据
| | |- ...
| |
| |- experiments/ # 本论文所述各类实验所用训练/验证/测试数据的元数据
| |- baseline/
| | |- train_all.csv
| | |- dev_all.csv
| | |- test_all.csv
| | |- ...
| |
| |- cross_generator/
| | |- train_tts.csv
| | |- dev_tts.csv
| | |- test_tts.csv
| | |- ...
| |
| |- cross_lingual/
| | |- train.csv
| | |- dev.csv
| | |- test_en.csv
| | |- ...
| |
| |- cross_speaker/
| |- train.csv
| |- test_same_spk.csv
| |- test_diff_spk.csv
| |- ...
|
|- LICENSE.txt
|- README.md

## 2. 音频说明

所有音频文件均采用WAV格式存储，采样率为16 kHz。

- `BD/`：包含英语与汉语的双语语音深度伪造音频及真实音频。
- `MD/`：包含46种语言的多语言语音深度伪造音频。
- `Real/`：包含源自LibriTTS、VCTK、AISHELL1、AISHELL3以及CommonVoice的真实语音数据。

如需获取详细信息与文件列表，请参阅`metadata/`目录。

## 3. 元数据格式

所有元数据文件均采用CSV格式，包含以下字段：
1. `file`：音频文件的相对路径
2. `label`：标注为`bonafide`（真实音频）或`spoof`（伪造音频）
3. `generator`：类型为`TTS`（文本转语音）、`VC`（语音转换）或`NV`（神经声码器）
4. `model`：语音生成模型的名称
5. `speaker`：说话人身份
6. `language`：语言代码（例如`en`、`zh`、`es`等）

关于元数据的目录结构：
- `metadata/BD` 与 `metadata/MD` 包含各语音生成器的元数据。
- `metadata/Real` 包含各真实数据集的元数据。
- `metadata/experiments` 包含论文中所述各类实验所用的训练、验证与测试划分的元数据，具体包括：
- 基线实验
- 跨生成器实验
- 跨语言实验
- 跨说话人实验

## 4. 许可协议

语音伪造（SpeechFake）数据集采用CC-BY-4.0许可协议发布。

如需获取完整条款，请阅读`LICENSE.txt`文件。

## 5. 引用方式

若您在研究工作中使用本数据集，请引用以下论文：

@inproceedings{huang2025speechfake,
title={SpeechFake: A Large-Scale Multilingual Speech Deepfake Dataset Incorporating Cutting-Edge Generation Methods},
author={Huang, Wen and Gu, Yanmei and Wang, Zhiming and Zhu, Huijia and Qian, Yanmin},
booktitle={Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)},
pages={9985--9998},
year={2025}
}