EmoTraceNet: Leveraging Residual Emotional Traces for Robust Audio Deepfake Detection

The rapid advancement of artificial intelligence and generative models has enabled the creation of highly realistic synthetic speech, giving rise to deceptive audio deepfakes. Specifically, Emotional Voice Conversion (EVC) poses severe security risks by manipulating the emotional prosody of an utterance to bypass psychological defenses while strictly preserving linguistic content and speaker identity. Existing Audio Deepfake Detection (ADD) systems exhibit significant vulnerabilities against EVC attacks, primarily due to extreme emotional imbalances in current public datasets and an inability to decouple emotional features from spoofing artifacts. To address these limitations, we first introduce DeepEmo, a large-scale, emotionally balanced audio dataset encompassing the latest EVC generative models. Furthermore, our empirical analysis reveals that current EVC algorithms struggle to entirely disentangle source emotions, leaving identifiable emotional traces. Motivated by this emotion trace hypothesis, we propose EmoTraceNet, a novel dual-branch framework integrating a spoofing detection branch and an auxiliary emotion classification branch. By employing a cross-task attention fusion mechanism, EmoTraceNet explicitly captures the residual emotional traces as well as generative artifacts. Extensive experiments on the DeepEmo and EmoFake benchmarks demonstrate that our proposed approach achieves state-of-the-art performance (3.06% EER on DeepEmo). Notably, EmoTraceNet exhibits superior cross-domain generalization, robust cross-emotion detection capabilities, and strong resilience against acoustic noise compared to existing baselines.