I get that record sessions are a huge hassle and simply paying VAs per AI-generated voice line is easier for everyone, but it somehow makes Paradox look a little careless to me.
Stories like these also set a precident. This is what voice ‘acting’ will be like for a moment before it becomes effectively eliminated because voice libraries will become diverse enough quickly and there will be no need for a single more voice actor to be included. It seems like VAs are basically forced to sell their voice to AI companies quickly to at least make a quick buck before they never get a job again.
There’s probably no stopping it, but that made this read all the more frustrating to me.
This is what (modern) voice acting has always been.
Actually read a few interviews with professional VAs or watch their streams if they do that. Two VAs actually interacting with each other and reacting is almost unheard of outside of very specific productions (and mostly are done as a stunt for some BTS footage). They read a dozen different takes of every line and go through like five different scripts worth of dialogue. And then they do “efforts” that are just general grunts and emoting that are used for the moment to moment gameplay and to pad out a scene that had heavy rewrites. It is why so many professional VAs can stream “their” games… because they genuinely have no idea what is going to happen.
Paying to train a limited use model off of a specific VA (or even a group of VAs) is the “logical” extension of that. And, arguably, it is a “good” one (with some MASSIVE caveats). Everyone lost their god damned mind over that FPS that came out last year where the announcer was (allegedly?) a model trained off of a VA. But it also meant that you could have stuff you would never have had otherwise. Nolan North isn’t going to get a paycheck to sit in a booth all day commenting on random matches. But a model that can read out a team’s name and string together different reactions? That is actually really cool and WAY better than the traditional sports game approach of “The Champion! just went through… A Table!”*
Like almost everything AI? The key is to focus on creators’ rights and control what can and can’t be used as training data. Because the genie is out of the bottle and ain’t going back in. But if we can protect the rights of what goes into training data? Then people are still paid for their effort/creation.
Do I think this was done “ethically”? I don’t know. But with everything Paradox has done in the past few years? I assume “not in the slightest”. But the concept is sound and one that we need to standardize sooner than later.
Of course, we also need UBI so that people’s lives aren’t tied to their jobs but that is a bigger mess.
*: Also, if you don’t think those aren’t already stitched and blended together with most of the same tech then I have a bridge to sell you
I’ll also add on that there are very good reasons to pay for models based on VAs. Brendan Fraser infamously permanently-ish hurt his vocal cords because of the performance that were expected of him in his prime. Same with a lot of VAs (I think David Hayter is one?) who basically need to smoke a pack a day when they are “in character” to get the right gravely voice. And while Stephanie Beatriz played it smart and made sure her “Rosa” voice was something she could maintain, a lot of actors and actresses basically can’t be the character they are famous for because it is killing them.
And pulling a solution out of my ass that is surely missing important aspects of the industry?
if I just want Nolan North or Felicia Day to voice a character then I buy the use of their model from their agency and am charged based on how much dialogue they have in a given game. If I want to use them as a character going forward (so what ANet tried with Felicia before they realized she was too expensive and decided to give Zojja permanent brain damage so she wouldn’t ever have dialogue again)? I can pay by line at a much cheaper rate.
But if I want Nolan North to do a voice that isn’t just Drake? Then I am paying him to train a new model and it gets a lot more expensive. And I can pay more to “own” that training data with the same caveats regarding future use. The main idea being that I want to make sure my Nolan North performance doesn’t end up in a competitor’s game next week.
I get that record sessions are a huge hassle and simply paying VAs per AI-generated voice line is easier for everyone, but it somehow makes Paradox look a little careless to me.
Stories like these also set a precident. This is what voice ‘acting’ will be like for a moment before it becomes effectively eliminated because voice libraries will become diverse enough quickly and there will be no need for a single more voice actor to be included. It seems like VAs are basically forced to sell their voice to AI companies quickly to at least make a quick buck before they never get a job again.
There’s probably no stopping it, but that made this read all the more frustrating to me.
This is what (modern) voice acting has always been.
Actually read a few interviews with professional VAs or watch their streams if they do that. Two VAs actually interacting with each other and reacting is almost unheard of outside of very specific productions (and mostly are done as a stunt for some BTS footage). They read a dozen different takes of every line and go through like five different scripts worth of dialogue. And then they do “efforts” that are just general grunts and emoting that are used for the moment to moment gameplay and to pad out a scene that had heavy rewrites. It is why so many professional VAs can stream “their” games… because they genuinely have no idea what is going to happen.
Paying to train a limited use model off of a specific VA (or even a group of VAs) is the “logical” extension of that. And, arguably, it is a “good” one (with some MASSIVE caveats). Everyone lost their god damned mind over that FPS that came out last year where the announcer was (allegedly?) a model trained off of a VA. But it also meant that you could have stuff you would never have had otherwise. Nolan North isn’t going to get a paycheck to sit in a booth all day commenting on random matches. But a model that can read out a team’s name and string together different reactions? That is actually really cool and WAY better than the traditional sports game approach of “The Champion! just went through… A Table!”*
Like almost everything AI? The key is to focus on creators’ rights and control what can and can’t be used as training data. Because the genie is out of the bottle and ain’t going back in. But if we can protect the rights of what goes into training data? Then people are still paid for their effort/creation.
Do I think this was done “ethically”? I don’t know. But with everything Paradox has done in the past few years? I assume “not in the slightest”. But the concept is sound and one that we need to standardize sooner than later.
Of course, we also need UBI so that people’s lives aren’t tied to their jobs but that is a bigger mess.
*: Also, if you don’t think those aren’t already stitched and blended together with most of the same tech then I have a bridge to sell you
I’ll also add on that there are very good reasons to pay for models based on VAs. Brendan Fraser infamously permanently-ish hurt his vocal cords because of the performance that were expected of him in his prime. Same with a lot of VAs (I think David Hayter is one?) who basically need to smoke a pack a day when they are “in character” to get the right gravely voice. And while Stephanie Beatriz played it smart and made sure her “Rosa” voice was something she could maintain, a lot of actors and actresses basically can’t be the character they are famous for because it is killing them.
And pulling a solution out of my ass that is surely missing important aspects of the industry?
if I just want Nolan North or Felicia Day to voice a character then I buy the use of their model from their agency and am charged based on how much dialogue they have in a given game. If I want to use them as a character going forward (so what ANet tried with Felicia before they realized she was too expensive and decided to give Zojja permanent brain damage so she wouldn’t ever have dialogue again)? I can pay by line at a much cheaper rate.
But if I want Nolan North to do a voice that isn’t just Drake? Then I am paying him to train a new model and it gets a lot more expensive. And I can pay more to “own” that training data with the same caveats regarding future use. The main idea being that I want to make sure my Nolan North performance doesn’t end up in a competitor’s game next week.