In-the-Wild: A Deepfake Detection Dataset

Welcome to In-the-Wild, a dataset for evaluationg audio deepfake detection.
It accompanies the paper: Does Audio Deepfake Detection Generalize? arXiv:2203.16263

---

Dataset Summary

The In-the-Wild dataset contains real and synthetic speech recordings of 58 celebrities and politicians, collected from online videos.

It provides a realistic benchmark for testing how well audio deepfake detection models generalize beyond laboratory data such as ASVspoof.

• Task: Audio Classification (Deepfake / Genuine)
• Languages: English
• Modality: Audio
• Size: 37.9 hours total
  • 17.2 hours fake
  • 20.7 hours real

---

Download

You can download the full dataset as a single ZIP file directly from this repository or via the Hugging Face datasets library.

Option 1: With the datasets library

from datasets import load_dataset

ds = load_dataset("mueller91/In-The-Wild")

Option 2: wget

wget https://huggingface.co/datasets/mueller91/In-The-Wild/resolve/main/release_in_the_wild.zip
unzip release_in_the_wild.zip

Citation

@article{muller2022does,
  title={Does audio deepfake detection generalize?},
  author={M{\"u}ller, Nicolas M and Czempin, Pavel and Dieckmann, Franziska and Froghyar, Adam and B{\"o}ttinger, Konstantin},
  journal={arXiv preprint arXiv:2203.16263},
  year={2022}
}