New phoneme-level framework explains why speech deepfakes are detected

As speech deepfake detection models grow more accurate thanks to self-supervised representations like wav2vec 2.0 and HuBERT, a key challenge remains: explaining why a sample is classified as genuine or spoofed. A new paper titled "Why Do You Say It Like That?" introduces a phoneme-level analysis framework designed to bridge this interpretability gap. The method is a post-hoc explainability technique applicable to convolutional neural network-based detectors. It leverages Gradient-weighted Class Activation Mapping (Grad-CAM) in tandem with a speech recognition system to produce saliency maps aligned with individual phonemes and pauses. This pipeline identifies statistically significant phonetic cues associated with different types of deepfake attacks and speaker characteristics, translating model decisions into linguistically meaningful terms.

Experiments conducted on the ASVspoof 5 dataset demonstrate that the framework maintains detection performance comparable to existing architectures while offering unprecedented interpretability. The authors show that the phonetic cues vary across spoofing methods and speakers, providing granular insight into how deepfakes differ from bona fide speech at the phoneme level. This work, authored by researchers from multiple institutions, addresses the growing demand for trustworthy and transparent AI in audio forensics. By connecting model predictions to measurable phonetic units, it enables human analysts to understand and trust deepfake detection outputs, which is critical for applications in journalism, law enforcement, and identity verification.