Guides and research on detecting AI-generated speech — cloned voices, deepfake audio, scam calls and synthetic narration. Written by the people building the detector.
The same recording can come back scored differently depending on which copy you upload. That is not the detector being unreliable, and understanding why makes you much better at using one.
As of 2 August 2026 this stopped being purely an ethics question in the EU. Here is where the line actually falls between a production tool and a thing you must label.
You run the same clip through three tools and get 91 percent AI, 34 percent AI, and inconclusive. Nothing is broken. Here is what those numbers actually are.
A detector score is the last thing to reach for, not the first. Here is the order of work that holds up when a newsroom has 40 minutes and a clip nobody can source.
The honest answer is that phone audio destroys most of the evidence a detector needs, and the reasons why are worth understanding before you buy anything that claims otherwise.
The known cases did not succeed because the audio was flawless. They succeeded because the payment process had a hole in it that a familiar voice could walk through.
A detector can tell you speech was synthesised. It cannot honestly tell you which product made it, and this is why that distinction matters more than it sounds.
Synthetic speech doesn't sound robotic anymore, so detectors stopped listening for that years ago. Here is what they actually measure, and where the measurement breaks.
Spot AI music from Suno and Udio in 2026 using audio, lyric, and metadata tells. Learn how spectral detection works and its honest limits.