Rethinking Synthetic Mixing for Real-World Scenefake Detection

Audio deepfake detection has largely focused on speech-level manipulations such as text-to-speech and voice conversion. In contrast, acoustic scene manipulation, known as SceneFake, alters the surrounding environment while preserving the speech signal, thereby changing contextual meaning without modifying speaker identity. For example, adding traffic noise to an indoor recording can falsely suggest an outdoor setting. These manipulations target scene-level consistency rather than speech content, making them difficult for existing audio deepfake detectors and often requiring manual, expert-driven analysis. Existing SceneFake benchmarks simulate such manipulations by mixing clean speech with mismatched background audio, implicitly assuming independence between speech and environment. However, real-world recordings violate this assumption: speech and acoustic context are inherently coupled through factors such as the Lombard effect and recording conditions. Consequently, synthetic mixtures fail to accurately reflect real acoustic scenes. This motivates the central question: Is synthetic mixing sufficient to simulate real-world acoustic recordings?