How are voices faked?
The principle is the same: You feed lots of recordings of the person you want to fake into an AI program, which chops up sounds and words into discreet bits; software can then rearrange the sounds so the subject can say anything you like. A team of sound engineers recently used deep-learning software to analyze 831 of John F. Kennedy’s speeches, and then created a convincing approximation of the 35th president reading the speech he was due to deliver the day he was assassinated. Researchers at the University of Washington last year synthesized realistic videos of Barack Obama speaking by mapping audio from one speech onto an existing video of him talking.