Voice cloning has moved past matching timbre. Most of the work now is in what happens between the words.

@HeyGen 's Professional Voice Clone trains on 30 minutes to 2 hours of your recorded audio and is built to keep breaths, pauses, laughs, and accent, along with how your vowels sound.

That data range is the point. Two hours of someone talking contains hundreds of hesitations, recoveries, half-laughs, and changes in pace. A short reference clip can’t show a model how you breathe before a long sentence or where your accent comes through when you speed up.

Those small irregularities are what listeners use to decide whether a voice belongs to a person.

HeyGen says the result is a voice no one can tell apart from yours.