Complex-valued neural networks for voice anti-spoofing

Phase information matters for detecting fake voices: The research demonstrates that audio phase information—typically discarded by traditional methods—is crucial for identifying spoofed or deepfaked voices. Ablation studies confirmed the model specifically learned to use this phase data, suggesting current systems may be missing important signals.

Explainability is achievable without sacrificing performance: Unlike raw audio-based models that are difficult to interpret, this complex-valued neural network approach maintains explainability while actually outperforming previous methods on the "In-the-Wild" anti-spoofing dataset, showing that transparency and accuracy aren't necessarily trade-offs.

Better audio representation preserves naturalness cues: By retaining both magnitude and phase information through complex-valued processing of CQT representations, the model captures more complete audio characteristics. This addresses a key limitation of magnitude-only spectrograms, which lose information about what makes audio sound natural versus artificially generated.

Technical improvements enable better deepfake detection: The research shows that thoughtful choices about how audio is represented and processed—combining frequency-domain analysis with phase preservation—can meaningfully improve detection systems' ability to identify voice spoofing and deepfakes in real-world conditions.