f233053/deepfake_detection_dataset_urdu

[](https://huggingface.co/datasets/CSALT/deepfake_detection_dataset_urdu)

# Deepfake Defense: Constructing and Evaluating a Specialized Urdu Deepfake Audio Dataset

This repository contains the Urdu Deepfake Audio Dataset introduced in the ACL 2024 paper "Deepfake Defense: Constructing and Evaluating a Specialized Urdu Deepfake Audio Dataset".

The dataset focuses on two spoofing attacks – Tacotron and VITS TTS – and includes bonafide audio samples for comparison. The dataset construction ensures phonemic cover and balance, making it suitable for training deepfake detection models in Urdu.

### Dataset Statistics

The dataset includes the following four parts:

1. Bonafide Part 1
2. Bonafide Part 2
3. Tacotron
4. VITS TTS

The statistics for each part are as follows:

| **Metric** | **Bonafide Part 1** | **Bonafide Part 2** | **Tacotron** | **VITS TTS** |
|------------------------------|---------------------|---------------------|--------------|--------------|
| **Total Duration (mins)** | 1,302.66 | 1,271.65 | 1,061.96 | 1,340.79 |
| **Max Sample Length (mins)** | 112.42 | 120.75 | 80.34 | 111.01 |
| **Min Sample Length (mins)** | 61.73 | 56.45 | 44.64 | 65.53 |
| **Avg Sample Length (mins)** | 76.63 | 74.80 | 62.47 | 78.87 |
| **Files per Speaker** | 708 audio files | 495 audio files | 495 audio files | 495 audio files |

## Structure

The dataset is organized into folders, each containing audio files for the respective parts mentioned above. Each folder is named according to its part (e.g., `Bonafide_Part1`, `Tacotron`, etc.).

## Usage

The dataset is available on Huggingface through the following link:
- Huggingface Dataset: https://huggingface.co/datasets/CSALT/deepfake_detection_dataset_urdu
The code for this project is on Github:
- https://github.com/CSALT-LUMS/urdu-deepfake-dataset

## Citation
```
@inproceedings{sheza-etal-2024-deepfake,
title = "Deepfake Defense: Constructing and Evaluating a Specialized Urdu Deepfake Audio Dataset",
author = "Sheza Munir, Wassay Sajjad, Mukeet Raza, Emaan Mujahid Abbas, Abdul Hameed Azeemi, Ihsan Ayyub Qazi, and Agha Ali Raza",
booktitle = "Findings of the Association for Computational Linguistics: ACL 2024",
year = "2024",
publisher = "Association for Computational Linguistics",
}
```

## Legal

CC BY-NC 4.0 license for the data hosted on HuggingFace and Google Drive.

[](https://huggingface.co/datasets/CSALT/deepfake_detection_dataset_urdu)

# 深度造假防御：构建与评估专用乌尔都语深度造假音频数据集

本仓库包含发表于ACL 2024会议论文《Deepfake Defense: Constructing and Evaluating a Specialized Urdu Deepfake Audio Dataset》中提出的乌尔都语深度造假音频数据集（Urdu Deepfake Audio Dataset）。

本数据集聚焦两类语音合成欺骗攻击——Tacotron与VITS TTS，并包含用于对比的真实（bonafide）音频样本。数据集的构建兼顾了音位覆盖度与平衡性，适用于乌尔都语深度造假检测模型的训练。

### 数据集统计信息

本数据集包含以下四个子集：
1. 真实样本集1（Bonafide Part 1）
2. 真实样本集2（Bonafide Part 2）
3. Tacotron生成样本集
4. VITS TTS生成样本集

各子集的统计信息如下：

| **指标** | **真实样本集1** | **真实样本集2** | **Tacotron生成集** | **VITS TTS生成集** |
|------------------------------|---------------------|---------------------|--------------|--------------|
| **总时长（分钟）** | 1,302.66 | 1,271.65 | 1,061.96 | 1,340.79 |
| **单样本最长时长（分钟）** | 112.42 | 120.75 | 80.34 | 111.01 |
| **单样本最短时长（分钟）** | 61.73 | 56.45 | 44.64 | 65.53 |
| **单样本平均时长（分钟）** | 76.63 | 74.80 | 62.47 | 78.87 |
| **每位说话者的文件数** | 708个音频文件 | 495个音频文件 | 495个音频文件 | 495个音频文件 |

## 数据集组织结构

本数据集按文件夹进行组织，每个文件夹对应上述子集，文件夹名称与子集名称完全一致（例如`Bonafide_Part1`、`Tacotron`等）。

## 使用方式

本数据集可通过以下链接在Hugging Face平台获取：
- Hugging Face数据集：https://huggingface.co/datasets/CSALT/deepfake_detection_dataset_urdu
本项目的代码托管于GitHub平台：
- https://github.com/CSALT-LUMS/urdu-deepfake-dataset

## 引用

@inproceedings{sheza-etal-2024-deepfake,
title = "Deepfake Defense: Constructing and Evaluating a Specialized Urdu Deepfake Audio Dataset",
author = "Sheza Munir, Wassay Sajjad, Mukeet Raza, Emaan Mujahid Abbas, Abdul Hameed Azeemi, Ihsan Ayyub Qazi, and Agha Ali Raza",
booktitle = "Findings of the Association for Computational Linguistics: ACL 2024",
year = "2024",
publisher = "Association for Computational Linguistics",
}

## 法律声明
托管于Hugging Face与Google Drive的数据集采用CC BY-NC 4.0许可协议。