ReplayDF: Audio Deepfake Replay Attacks with 109 Hardware Combinations

ReplayDF contains between 100,000 and 1,000,000 audio samples designed to evaluate the impact of physical replay attacks on deepfake detection systems. Created by mueller91 and updated in 2025, it features re-recorded bona-fide and synthetic speech across six languages and 109 speaker-microphone combinations.

Use Cases

• Evaluating the performance of audio classification models against physical replay attacks
• Benchmarking deepfake detectors using the 109 speaker-microphone combinations
• Analyzing the impact of diverse acoustic environments on spoofing detection accuracy

Strengths

• Includes 109 unique speaker-microphone hardware combinations
• Covers six different languages including English, Italian, and Polish
• Features speech generated from four distinct TTS models

Limitations

• Restricted to non-commercial use under CC BY-NC 4.0 license
• Relies on source material from M-AILABS and MLAAD v5 rather than original primary speech recordings

Provenance

Source

mueller91 (Hugging Face), based on M-AILABS and MLAAD v5 datasets

Collection Method

Re-recording bona-fide and synthetic speech in diverse acoustic environments using various hardware

Freshness

Last updated June 2025

Geography

Multilingual (6 European languages)