You aren’t “[creating] a novel voice from scratch, that’s just not how the technology works. It needs a human to extract data from and compile something intelligible.

Much like with art AIs, the outputs don't necessarily have to slavishly mimic the style of any of the inputs. Train an AI with a bunch of different voices and then you can get it to generate a novel voice that isn't a copy of any specific one that it was trained on.

Using computer generation to imitate a person using their own biometric data should be illegal unless explicit consent is given.

This doesn't affect what I've said. If imitating a specific human comes with a bunch of annoying legal and economic hassles, then don't imitate a specific human. Create a novel voice and you're free of all of that.

And yes, the technology lets you create a novel voice different from any of the ones it was trained on. I do know how these things work.