The actual Sky voice actor was interviewed and, surprise, she sounds way more like the OpenAI voice than Scarlett.
That said, this is certainly a problem we will see in the future.
That said, this is certainly a problem we will see in the future.
Copying a comment from another HN user (not sure if appropriate to name them):
* This one: https://www.youtube.com/watch?v=vgYi3Wr7v_g
* compare it to: https://youtu.be/GV01B5kVsC0?feature=shared&t=158
That seems like quite a strong resemblance to me.
I wanted to laugh at and belittle openai when I first saw scarjos claim that her friends and family couldn't tell the difference but from the samples you just linked I find that so unlikely that I think she is lying.
> It looks? like you're part of some recording? or production setup? With those lights? tripod? and possibly a mic? It looks like you might be gearing up to shoot a video? or maybe even a live stream?
They both have a vocal fry, but the former sounds more like valley girl without the slang or vowel lengthening.
They still created a voice that sounds like ScarJo - and many signs point to that being intentional. Overall, not unlike the case in "Tom Waits vs. Frito-Lay, Inc".
The Waits case is highly dissimilar, IMO. But tbf this isn't an area of law I'm expert in, and my knowledge of Right of Publicity law is mostly from law school, which was... a while back.
A voice isn't that unique in the grand scheme of things.
https://www.vice.com/en/article/wxepzw/musicians-algorithmic...
The problem OpenAI has is they tried to get Scarlet to do the voice and after she declined OpenAI found someone who has a similar voice to do it. That shows intent which would otherwise be difficult, combined with the actual result being closer to Scarlet’s voice than Sky is naturally and they are likely going to lose.
Edit: By after I am saying they decided to use Sky’s voice not that it was recorded first.
As the public currently knows based on reporting from WaPo, the original VA was selected and performed her role (months) BEFORE they reached out to Scarlett for the first time. Please don't misconstrue the timeline for other people that are just now tuning in.
The core question eventually put before a jury would be if it was unintentionally similar or an imitation and attempting to pay her for use of her voice is clear evidence they wanted it because of it or they perceived the voices where overly similar. Either of those are possible, but to then use the voice anyway became problematic.
I’m not saying Sky did anything wrong. But the final result is close enough many people assumed it was Scarlet without promoting, that’s problematic when they are making references to her role as an AI voice and obviously wanted to make a deal.
And of course, some specific aspects of the case are unique. Universal used a face mold they had made of Glover's face, for instance.
I’ll give you the mold makes the intent more clear, but I doubt the jury is going to be debating intent. Pulling the voice suggests OpenAI thinks this is either a real risk or bad publicity.
Really using a voice on a digital assistant that sounds like the actress playing a role of a digital assistant is just dumb. Especially if you then tweet about the film days before releasing the voice.
But it’s a small number of voices who all sound distinct. So it seems unlikely they wanted to released two similar voices in that context. Especially when it specifically sounds like the voice actor in a film as an AI.
If they had released the voice they could have argued they wanted something different, but by waiting and going with Sky after the negotiations failed it’s a different matter.
they tried to get Scarlet to do the voice and after she declined OpenAI found someone
This is exactly the question the Washington Post set out to answer, and the answer is "no": that didn't happen.
Simply recording Sky’s voice beforehand doesn’t imply a different intent. This isn’t animation where you’re trying to match two mediums and you want the actors recording first. Having a similar voice to work with would have been useful internally to know if they wanted to pay what it might cost to use her voice.
Here’s what probably happened. OpenAI people realize they can do TTS for ChatGPT so they hire some voice actors. Some people point out that it would be cool if it could be as good as it is in Her, especially with 4o on the way, so they decide to contact Scarlett Johansson either to do a voice or just some promotional thing. She declines, but after 4o is released OpenAI realizes they can fairly make the comparison to the movie based on the quality and speed of the voice. Simply subtly referencing the film via an obvious parallel doesn’t constitute deceptive impersonation.
Personally there’s many ways they could have been less suspicious from my perspective. Have a large number of voices or the ability to tweet them in many ways and no one voice seems important. Contact many famous people and it would have seemed like a promotional rather than a reference. Have someone lower on the totem pole than the CEO of a multi billionaire dollar corporation trying to make the deal. This not being about an actor playing an AI assistant in a movie vs the voice of an AI product. Not actually making the reference in a tweet.
The phrase ‘the preponderance of the evidence’ comes to mind.
There’s a chance they did contact other celebrities. We only know for sure they contacted Scarlett due to her lawsuit.
Here’s the thing. No matter what, people would make the reference to Her. It’s like how everyone would make the connection to any sci fi movie if a device came out similar to the one in it. A woman’s voice as a virtual assistant is a very broad UI category, and being cheeky and drawing the comparison to a specific movie on everyone’s mind doesn’t constitute misleading impersonation, especially when you consider from OpenAI’s view the fact that they hired a different actress and didn’t ask her to do a Her impersonation.
> especially when you consider from OpenAI’s view the fact that they hired a different actress and didn’t ask her to do a Her impersonation.
That’s not specifically what they are accused of doing. The voice actress’s raw audio is only half the story, how they manipulated that audio makes a real difference in what the model sounded like. As many people assumed it was her objectively the end result was close. Worse the model sounds closer to Scarlet than the voice actress does.
Put another way, they wanted to end up at X and also happened to end up close to X when the search space was huge. Where did you get that Huge TV? It’s the craziest thing it just happened to fall off a truck.
https://www.washingtonpost.com/technology/2024/05/22/openai-...