Source: Arstechnica, Nov 2016
Adobe has demonstrated tech that lets you edit recorded speech so that you can alter what that person said or create an entirely new sentence from their voice. It seems inevitable that it will eventually be referred to as “photoshop but for audio.”
The tech, dubbed VoCo (voice conversion), presents the user with a text box. Initially the text box shows the spoken content of the audio clip. You can then move the words around, delete fragments, or type in entirely new words. When you type in a new word, there’s a small pause while the word is constructed—then you can press play and listen to the new clip.
VoCo works by ingesting a large amount of voice data (about 20 minutes right now, but that’ll be improved), breaking it down into phonemes (each of the distinct sounds that make up a spoken language), and then attempting to create a voice model of the speaker—presumably stuff like cadence, stresses, quirks, etc., but Adobe hasn’t provided much detail yet.
Then, when you edit someone’s speech, VoCo either finds that word somewhere within the 20-minute clip—or if the word hasn’t been uttered, it is constructed out of raw phonemes. At around the 4:30 mark in the video you can hear the phase “three times” being constructed from scratch; if you listen carefully, it does sound a bit synthetic, but it’s not awful. Copying and pasting existing words sounds better.