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.