SpeechAntiSpoofingBenchmarks/InTheWild

In-the-Wild Audio Deepfake Dataset

Benchmark-ready packaging of the In-the-Wild audio deepfake dataset for speech anti-spoofing / synthetic-voice detection.

Overview

In-the-Wild (Müller et al., Does Audio Deepfake Detection Generalize?, arXiv 2203.16263) pairs genuine speech with audio deepfakes of politicians and public figures, collected from publicly available sources. It is a cross-domain generalization benchmark: models trained on lab datasets (e.g. ASVspoof) are evaluated here against real-world conditions. The task is binary classification: bonafide (genuine human speech) vs. spoof (deepfake). 31,779 clips (19,963 bonafide / 11,816 spoof), 16 kHz mono.

License & redistribution

Redistributed under the Apache License 2.0; the full text is in LICENSE.txt. Audio is the original 16 kHz mono signal encoded to FLAC (16-bit PCM). We thank 'VocalSynthesis' for the audio deepfakes included in the source dataset.

Schema

Canonical 4-column parquet: path (string), audio (Audio(16000)), label (ClassLabel[bonafide, spoof]), notes (JSON string with a unique utterance_id, the speaker name, and the source label string).

Citation

@inproceedings{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},
  booktitle={Interspeech},
  year={2022}
}