Audio deepfake detectors have a labeling problem they were not built to solve.
Researchers Xiaochao Zou, Shuai Wang, and colleagues — citing arXiv:2603.14033 — trained detection systems on a fundamental assumption: a voice clip is either real or fake. But when a genuine recording is run through speech restoration or voice quality conversion, that binary breaks. The underlying speaker and what they said remain intact, yet the audio has been processed. Current detectors, trained on SSL representations and tested via DF-Arena fine-tuning, can usually spot that something was done to the audio — but they struggle to tell whether the original source was real or synthetic. Processed genuine speech and processed synthetic speech look too similar from the outside.
The gap matters because speech restoration tools are increasingly mainstream. Podcasters clean up recordings, call centers scrub background noise, and accessibility tools smooth out speech impediments. Any of these workflows can now cause a legitimate voice clip to trip a deepfake detector — or, conversely, let a doctored one pass because its tell-tale artifacts were laundered away by a restoration step. The researchers argue that defenses need to report three things separately: source authenticity, processing status, and where in the clip any manipulation occurred.
It is a reasonable ask, but it also raises the bar considerably for a detection field that has spent years optimizing for the simpler binary problem — and that is already struggling to keep pace with synthesis quality.