What Do Deepfake Speech Detectors Actually Hear?

2026-06-09

AI Summary

Current deepfake audio detectors lack interpretability, making it difficult to understand their decision rationale. This work proposes an audio-native explainable method that leverages WavLM self-supervised representations and employs Integrated Gradients to localize critical evidence in the time domain. The attribution results are validated through causal masking and human semantic annotations. For the first time, this approach provides a fine-grained, verifiable time-domain explanation mechanism for fake speech detection, revealing that models with comparable performance—such as AASIST, CA-MHFA, and SLS—actually rely on markedly different semantic cues: environmental noise, phonetic artifacts, and word boundaries, respectively. These findings move beyond prior empirical speculation by offering concrete, interpretable insights into model behavior.

Abstract

Deepfake speech detectors often output a single score without explaining why an audio sample is flagged, where in the signal the evidence lies, or what cues drive the decision. We propose an audio-native explainability pipeline using Integrated Gradients on time-aligned self-supervised representations to localize decision evidence over time. We apply the proposed method to three WavLM-based detectors (AASIST, CA-MHFA, SLS) on ASVspoof 5 and manually annotate the highest-attribution regions to provide a semantic meaning of the most important cues. Despite similar performance, the detectors rely on different cues: AASIST emphasizes non-speech/environment cues, CA-MHFA focuses on localized phoneme artifacts, and SLS relies on word boundaries and spectral integrity. We move beyond speculative reasoning and validate our findings by causal masking of the primary detector cues. Observed performance degradation further supports the explained detector semantics.

Problem

• deepfake speech detection
• explainability
• decision evidence
• audio forensics
• self-supervised representations

Innovation

• explainable AI
• deepfake speech detection
• Integrated Gradients
• self-supervised representations
• causal validation