Researchers split deepfake audio detection into two stages — finding where the fake starts, then proving it's fake

Computer scientists built a new approach to detecting partial deepfakes — audio clips where only parts are artificially generated, not the whole thing. Instead of judging the entire clip at once, the system first finds where the fake begins and ends, then separately checks each section to determine if it's real or synthetic.

Until now, deepfake detectors worked like a smell test: sniff the whole thing and decide if it stinks. But partial deepfakes are harder — most of the clip is genuine, so the overall 'smell' stays normal. This method treats it as a localization problem first, an authenticity problem second, which means it can catch splices that older detectors miss. The practical effect: audio forgeries get harder to deploy without getting caught.