I do have Luddite concerns about the potential impact of AI like this on voice actors, both in terms of reducing the amount of work and the possibility of their performances being used as inputs for AI without credit or reward, but I can also see the huge potential advantages for modders and smaller independent games producers in particular, so as with most technological advances I’m sure there will be pros and cons. But if we’re so close to being able to have AI turn in creditable voice performances, sounds like those implications need to be worked through with some urgency! I assume they’re already hot topics of debate amongst interested parties.
As far as modding goes, it doesn't really surpass in my eyes what we already do with other people's work. As long as nobody tries to profit from it, or claim it as a wholly original work, I don't think the powers-that-be should treat it more seriously than anything else.
As for completely replacing the need for VO, I don't really have a good answer to that. As far as I know (which isn't a lot) "A.I." doesn't actually create anything new, all of human recorded history has been sitting in a great stew of that scientists poured into an algorithm, that through extensive use of flash cards and night classes, is able to determine what sounds correlate to what emotion, and what emotions to what context, and whatever else they could think of regarding linguistics. Right now I think it's still easier to use a person to accomplish those things, but like likeness rights to movie stars, it might not be too far off where voice actors sit down and read a few dozen prompts and get a check in the mail whenever their voice is used. I think James Earl Jones, already did something like to allow his voice to be used for Darth Vader in perpetuity.
Wow! I’ll confess it’s years and years since I tried Viconia’s romance in BG2 so can’t recall how much was already voiced, but that sounds really good. I’m a total AI (and voice synthesiser) ignoramous, so would be interested to know more about the process you went through.
The initial line of each LoveTalk is what originally existed in the game. I used ElevenLabs voice synthesizer because it was the first time a Voice AI actually sounded real to me. And for some reason Youtube thought I should be made aware of dozens of Dagoth Ur memes using it (not wrong). I extracted all the sound files from Baldur's Gate II, listened to all of Viconia's lines, separating them by tone, I extracted one of the dialogue files, and signed up with ElevenLabs. I created I think what eventually became 6 Viconia voices, by putting in lines only using a certain tone, I then started generating the lines with the dialogue used. You have to play around with the text to get the right read, add line breaks, periods, change a word here or there to something more phonetically appropriate. You generate the dialogue half a dozen times, until you get a few workable one, or sometimes I would splice a few together with Audacity. I did that for most of the longer lines, you can probably tell. Viconia sometimes has a thicker accent too, I'm not sure where that came from. You can see one of the lines I kept trying to get a better read of in the outtakes toward the end, it was like that but usually for longer lines, maybe I'll put up a video with more stuff like that. With the minimum subscription at ElevenLabs, you only get 30,000 characters a month to work with, so I couldn't spend too much time with every line I would have liked. Still, there are 2 or 3 in there I think are spot on, and more than a few that are passable.
