Voice analysis shows striking similarity between Scarlett Johansson and ChatGPT | Hacker News Reader

Voice analysis shows striking similarity between Scarlett Johansson and ChatGPT

npr.org

41 pointscroes26 comments

> The researchers measured Sky, based on audio from demos OpenAI delivered last week, against the voices of around 600 professional actresses. They found that Johansson's voice is more similar to Sky than 98% of the other actresses.

That feels like, how to lie with staristics

In a sample of only 600 voices, there were ~12 actresses whose voice was more similar to chatgpt than Johansson. Presumably many of these samples were significantly different (i assume it covers a wide range, e.g. gravely deep male voices, people with different accents). The fact there are 12 more similar matches is striking to me.

If anything, i feel like this is strong evidence for the opposite of the headline.

>covers a wide range, e.g. gravely deep male voices

It does say "600 professional actresses", so at least they have some limits to their lying with statistics. :-)

12 out of 600 is 2%, just like the headline says. I don't think the headline putting the number in relative terms is particularly misleading.

> I don't think the headline putting the number in relative terms is particularly misleading.

"Scarlet's bogus claim: here are 12 actresses who you might know who sound more like Sky than her."

Is how you would present the same information with the opposite slant.

"The analysis of Sky often rated Hathaway and Russell as being even more similar to the AI than Johansson" So how does that support the title?

The headline says "strikingly similar". I do not think that fits the numbers.

In the last 50 people you talked to, did 1 of them have a voice that struck you as strikingly similar to ScarJo?

So from the article : 2% of 600 professional actresses voices are more similar to Sky's than Scarlett Johansson's. This means than for 3 feminine voices provided by open AI, there was about 1 in 17 chances to get a voice that similar to Johansson's without planning it. That's not something that seems too unlikely.

Right; unfortunately the CEO tweeted

> there was about 1 in 17 chances to get a voice that similar to Johansson's without planning it

Now, add in the chances "without planning it" of a) reaching out to the actress to request permission to use her voice and b) tweeting that Her is your favorite movie.

>The researchers found that Sky was also reminiscent of other Hollywood stars, including Anne Hathaway and Keri Russell. The analysis of Sky often rated Hathaway and Russell as being even more similar to the AI than Johansson.

This gives the headline a slightly different meaning.

If anything this suggests openai trained Sky on more than one Hollywood actress voice.

No, this study alone doesn't suggest that. Rather, it suggests that a lot of people have indicernable voice

I wonder how things would have turned out if OpenAI never contacted Johansson or referenced the film 'Her' in a tweet.

I saw Her, I watched the GPT-4o keynote and personally didn't notice any connection in the voices until it made the press.

The problem I see ahead for OpenAI is not whether or not they actually used Johansson in the training¹, but rather that they made public reference to her and that is likely to run afoul with her right to publicity.

¹If the "Sky" voice and expression comes wholly from voice/data sources other than Johannson, this would be trivial to show.

Surely you'd do the Star Trek computer if you were real nerds rather than a single 2013 film. I can't even visualise (audiolise?) Johansson's voice, it's not particularly special to me. Yet I'd recognise Majel Barrett before end of the first word.

Though it would probably be more accurate if they used the autopilot voice from Wall-E but Apple would really stomp them for that.

So basically it's like they had chosen the most similar voice from 50 random actresses.

How do we know or how can one prove the voice was not trained on audio from her movies?

It would be hard (if not technically impossible) to prove it wasn’t trained on specific material. To prove it was, however, it’d be sufficient to show that the model can be made to produce training data ‘verbatim’. In this case, perhaps a certain sentence would be read in a way that is so similar to some piece of source material so as to make it very unlikely to not be the direct result of training on that data.

Burden should be in the other direction. Otherwise we end up with a Russel's teapot type situation.

Yes. Prosecution should perform exhaustive search thru company policies, internal communications and processes during discovery to get the the bottom of this.

Clickbait: there other famous actresses that are even more similar to this AI voice than Johansson's voice.

If you did a comparison with "random American woman" I'm betting you'd find even more similarities.

IMO, there's nothing particularly distinctive about Johansson's voice. She's certainly no Lauren Bacall, Tilda Swinton, or even Fran Drescher.

If nobody said anything, would anyone know or care?

What if they just thought her voice set the mood. They couldn't get Johansson so they set out to find something similar and found it.

So what? They didn't say it was Johansson so no one could claim she endorsed it. How many times have you heard something on the radio/TV and thought it was some actor when it wasn't?

If she said nothing in public, nothing would have come of this. She could have writen them in private.

This is much ado about nothing and nothing happened.

I think it’s PR. Don’t know for what or why but nobody I knew, or myself, had even thought of SJ when we saw the videos. Absolutely nobody cared who the voice sounded like the only thing that mattered was that it sounded like it was coming on to you. Very weird tbh

> They couldn't get Johansson so they set out to find something similar and found it.

> So what?

You can't just do that though. There's at least two cases of legal precedent where companies had to pay out millions to artists, though musicians in these cases, for soundalikes after they couldn't negotiate originals.

Midler v. Ford Motor Co, 1988 (won on appeal)

Waits v. Frito-Lay, 1992 (won and won appeal)

I won't comment further since I'm certainly not a legal expert and this is going to be settled by the parties or will be determined by a court. I mostly wanted to mention "Tom Waits sued Doritos"

These guys are now doing it for th views

Reply on news.ycombinator.com