Found a weird trick for spotting deepfake audio with a waterfall spectrogram

I was stuck on a dodgy voicemail from my boss's bank request, then tried pulling the audio into Audacity and switching to spectrogram view - the fake voice had these perfect blocky patterns that real speech never makes. Has anyone else tried visual analysis for voice scams, or am I just late to the party?

2 comments

2 Comments

logan7051mo ago

Yeah it's the same thing as those AI generated images where the hands look all wrong or the background text turns into gobbledygook. The tech is getting better but the artifacts are still there if you know where to look. I've been checking suspicious voicemails from "my bank" by looking for weird silences in the waveform, like the pauses between words are too uniform or the syllables lockstep into this perfect rhythm real people never use. Pretty soon these tells will be gone too but for now it's a solid edge.

bens811mo ago

Pull up the spectrogram view and look for those perfect little rectangles where the computer is stitching together individual phonemes. Real voices bleed and smear across the frequencies, but AI voices lock into these sharp, clean blocks like a digital stencil. I also check the silence between words in the waveform view, real people have uneven pauses from breathing and thinking, but the fakes have mechanical gaps that are all the same length. Another dead giveaway is the breath noise, or rather the lack of it. Real recordings have subtle background hiss and room tone that stays consistent, but these deepfakes often have dead silent gaps that are too clean. If you want to be thorough, flip it over to the pitch contour view and watch for notes that hold unnaturally steady like a robot singing one note.