Early-Reflection Wiener Filter and Spectral Subtraction for Replay-Resilient Audio Deepfake Detection

Abstract

With the rapid advancement of artificial intelligence (AI) and deep learning (DL), audio deepfakes have become increasingly realistic, posing significant challenges for automated detection systems. Among various attack types, replay attacks, where bona fide or synthetic speech is played through a loudspeaker and re-recorded via a microphone, introduce environmental and device-induced distortions that degrade detection performance. In this work, we analyze replay attack characteristics and propose a signal-level compensation pipeline to enhance detection robustness. The pipeline sequentially applies DC offset compensation, spectral subtraction, and an early-reflection Wiener filter (ERF) to mitigate device bias, additive noise, and channel-induced distortions. Experiments on the ReplayDF dataset (109 microphone–speaker pairs) show that the proposed pipeline improves signal-to-reverberation ratio (SRR) and reduces equal error rate (EER) across multiple state-of-the-art detectors. The proposed framework is evaluated under a measured or estimated RIR setting and demonstrates consistent detector-independent improvements under replay attack conditions.