The world is being quietly rearranged by people who write very long documents.


The title they went with Does Audio Deepfake Detection Generalize? Noisy translates that to

Audio deepfake detectors fail in the real world


Researchers tested audio deepfake detection systems on real celebrity and politician recordings and found they perform dramatically worse than laboratory benchmarks suggest — sometimes failing up to ten times more often. This reveals that the detection methods the field has been building work well only on the specific datasets researchers use to train them, but collapse when confronted with actual deepfakes in the wild.
The audio deepfake detection field has spent years optimizing solutions to pass internal benchmarks while missing the actual problem: detecting fakes that sound convincing to human ears in real recordings. This gap means voice cloning technology is advancing faster than our ability to identify it, which matters as synthetic audio becomes cheaper and easier to produce.

If you insist
Read the original →