ReplayDF

ReplayDF ReplayDF** is a dataset for evaluating the impact of **replay attacks** on audio deepfake detection systems. It features re-recorded bona-fide and synthetic speech derived from M-AILABS and MLAAD v5, using **109 unique speaker-microphone combinations** across six languages and four TTS models in diverse acoustic environments. This dataset reveals how such replays can significantly degrade the performance of state-of-the-art detectors. That is, audio deepfakes are detected much worse once they have been played over a loudspeaker and re-recorded via a microphone. It is provided for **non-commercial research** to support the development of **robust and generalizable** deepfake detection systems. 📄 Paper Replay Attacks Against Audio Deepfake Detection (Interspeech 2025) 🔽 Download 📌 Citation 📁 Folder Structure 📄 License Attribution-NonCommercial-ShareAlike 4.0 International: https://creativecommons.org/licenses/by-nc/4.0/ Resources Find the original resources (i.e. non-airgapped audio files) here: MLAAD dataset v5, https://deepfake-total.com/mlaad. M-AILABS dataset, https://github.com/imdatceleste/m-ailabs-dataset. Mic/Speaker Matrix 📊 Mean Opinion Scores (MOS) The scoring criteria for rating the audio files are outlined in the table below: | Rating | Description | Speech Quality | Distortion (background noise, overdrive, etc.)| |-----------|---------------|---------------------------|-----------------------------------------------| | 5 | Excellent | Clear | Imperceptible | | 4 | Good | Clear | Slightly perceptible, but not annoying | | 3 | Fair | Understandable | Perceptible and slightly annoying | | 2 | Poor | Understandable | Perceptible and annoying | | 1 | Very Poor | Barely understandable | Very annoying and objectionable | | e | Error | Inaudible | Heavy |