I'm not really up-to-date on voice synthesis. Have we reached the point where we can get enough training data from just a handful of voice actors to train a model of this quality?

Or is this a case of them using those voice actors for fine-tuning a pretrained model and just being quiet about that?

[–] Dremor@lemmy.world 6 points 6 months ago* (last edited 6 months ago) (6 children)

https://commonvoice.mozilla.org/

[–] Mirodir@discuss.tchncs.de 0 points 6 months ago (1 children)

Yeah, if Mozilla's goal is 1200 clips/day and 2400 validations/day then I have a strong suspicion that Stellaris uses a pretrained model and there are no royalties for the people whose voices were used for the pretraining. Not that it would be feasible to spread royalties among that many people in the first place.

What could point against that suspicion though is that Stellaris doesn't need a "perfect" model so maybe they can get away with much, much less. After all the whole gimmick is that it is in-universe AI. A (near-)flawless model would be (near-)indistinguishable from a regular voice actor. Then there would've been no need to hire a bunch of voice actors to train an AI in the first place.

Assuming that it is pretrained -> finetued though, the only hope is that those initial files were donated willingly and not scraped somewhere. Otherwise their "ethical" argument goes out the window.

[–] Amaltheamannen@lemmy.ml 4 points 6 months ago

They claimed they specifically used an ethical model with a license where they pay the person whos voice was trained on.

load more comments (4 replies)