I really wish they took actual things the individual said in the past, but used a deepfake for them actually saying it. Most of the deepfakes were easy to spot, but a couple weren't clear at first that they were fake, only to be given away by the content of their speech.
But the person saying something outrageous would undermine the point of the study--to determine if deep fakes were indistinguishable from actual footage. If it's given away by something unrelated to the deepfake method, then you're adding in an external bias unrelated to the realism of the deepfake.
Unless I misunderstood the goals of the study of course.
It doesn't necessarily have to be. For example what if you modified a clip of a political candidate asking for funding and only modified the name of the organization/website their supporters should go to in order to donate.
Getting the voice accurate would be very hard, even if you only used existing words they've said to make up the false domain. I've tried editing audio that way and it still sounds like when a video game ai stitches dialog together.
Getting the voice accurate is actually really easy, and is done all the time. https://en.wikipedia.org/wiki/Overdubbing is used in music and movie production, and you can change out a few words with "ease" [1]. You use an impressions actor, get them to say the whole line, matching the cadence, and you've got a decent amount of skew in mouth movement because people aren't always lipreading carefully.
None of these techniques are out of the budget of a college film or dedicated hobbyist.
Overdubbing is far easier with an instrument than with a voice, the article doesn't even list examples of an impressionist over dubbing someone, just overdubbed duets. Also what you're describing is not overdubbing, it's ADR. Overdubbing is just that, dubbing over the existing audio. Not only would you need to mimic the speech but you'd have to recreate the ambient sound as well. Even with a person speaking into a microphone in a silent room the change in ambience is noticeable unless the whole scene is using a false ambience track separate from the voice track.
And I disagree with you about the lipreading. There's a reason good ADR in films is always done with a shot of the back of the actors head.
Even when it's done with literally the same original speaker it's pretty obvious. Or have you never fooled a stranger in the alps?
