32
Serious question, how do you spot voice cloning in deep fakes now?
I used to just look at the mouth movements to catch lip sync errors, but last week my buddy sent me a voice clip that sounded exactly like our boss and I couldn't tell it was fake until I called him back. How do you guys check for audio deep fakes when the video looks perfect?
4 comments
Log in to join the discussion
Log In4 Comments
theas283d ago
Three seconds is actually the number I keep seeing in the research papers about this. I mean, it makes sense when you think about it, most of the current voice cloning tools need at least three seconds of clean audio to make a decent copy. The breathing thing mark_chen62 mentioned is spot on though, that's where these things fall apart every time. So if you hear a voicemail that's shorter than three seconds and it sounds a little too clean, that's actually probably the real person. But anything longer than that where the pauses are too perfect, yeah that's the red flag. The lunch question trick is good too, but I'd add that you should ask something really specific like "what color was my car when you saw it last week?" because clones can't handle details like that.
6
felixm2917d ago
Has anyone actually tested this with their own voice? The real trick is that voice cloning still struggles with irregular breathing patterns and those little throat clear sounds people make. @young.thomas brings up a good point about voicemails, but I'd say the bigger giveaway is when the cloned voice handles pauses weirdly. Real people hesitate and stumble a bit, but the AI versions tend to have this too-perfect rhythm that just feels off after a while. Another thing to check is asking for something specific from the person's own memory, like what they ate for lunch yesterday or something trivial. If they can't answer that right away, it's probably a clone trying to buy time.
5
mark_chen6217d ago
Tested it with my own voice last week. The breathing thing is real for sure. I noticed the AI version of me didn't do that little exhale I always do before answering a question. Also the pauses were too even, like a metronome. My buddy called me out on it immediately because it sounded "too clean.
4