by Dan Moren
Behind Apple’s new voice cloning feature
Fast Company’s Harry McCracken talked to some of the team behind Apple’s new Personal Voice accessibility feature about its development as well as some more fine details:
When it came to enabling third-party apps to speak via Personal Voice, Apple put privacy measures in place similar to those it imposes for photos, location, and other bits of personal data in its care. Such apps can only hook into Personal Voice with the user’s permission, must be running in the foreground, and receive only enough access to read text in the voice, not to get at the data used to generate it.
I’ve been thinking a lot about the privacy implications, but the implementation of this feature certainly seems that it will be harder to abuse than something like ElevenLabs’s voice cloning tech. For example, just having to spend fifteen minutes training the model with a random set of words is going to make it a lot harder to create a model of someone else’s voice without their knowledge, even if it does give me shades of training the ViaVoice dictation software circa 2000 by reading Treasure Island to it.