Hacker News Reader

Comment by bnralt | Hacker News Reader

Right. Even if you think OpenAI isn’t a good place, this is an investigation by an established newspaper that refuted some of the more serious accusations (that OpenAI got a Johannson impersonator - they didn’t, that they modified the voice to sound like Johansson - evidence suggests this didn’t happen). When the reaction is “I don’t care that an investigation refuted some of the accusations”, it demonstrates someone isn’t openly approaching things in good faith.

Likewise, if someone’s attitude is - “OK, maybe there’s no paper trail, but I’m sure this is what the people were thinking”, then you’ve made an accusation that simply can’t be refuted, no matter how much evidence gets presented.

View on HN

> refuted some of the more serious accusations (that OpenAI got a Johannson impersonator - they didn’t

A lot of the argument here comes down to whether the article does refute that. I don't believe it does.

What it refutes is the accusation that they hired someone who sounds like Johansson after she told them she would not do it herself. That was certainly a more damning accusation, but it's not an identical one.

But in my view, it requires a pretty absurd level of benefit of the doubt to think that they didn't set out to make a voice that sounds like the one from the movie.

Maybe good for them that they felt icky about it, and tried to get her for real instead, but she said no, and they didn't feel icky enough about it to change the plan.

Do you believe the article "refutes" that? Does it truly not strike you as a likely scenario, given what is known, both before and after this reporting?

> A lot of the argument here comes down to whether the article does refute that.

It clearly refutes the claims that they got a Johansson impersonator. The article says this is a voice actress, speaking in her normal voice, who wasn’t told to mimic Johansson at all. You can say that you personally think she was chosen because people thought she sounded similar to Johansson, even though there’s no evidence for that at this point. But the claim - which was made several times in discussions on here before - that she is a Johansson impersonator is factually incorrect.

> But in my view, it requires a pretty absurd level of benefit of the doubt to think that they didn't set out to make a voice that sounds like the one from the movie.

I tried it several times in the past and never once thought it sounded like Johansson. When this controversy came out I looked at videos of Her, because I thought Johansson could have been using a different voice in that movie, but no - the voice in her is immediately recognizable as Johannson’s. Some have said Sky’s was much closer to Rashida Jones, and I agree, though I don’t know how close.

I think this is quibbling over the definition of "impersonator"?

I think the most plausible thing that happened is that they thought "hey it would be so awesome to have an AI voice companion like the one in Her, and we can totally do that with these new models", and then auditioned and hired someone that sounded like that.

Does it not fit the definition of "impersonator", since they didn't explicitly tell the person the hired to impersonate the voice from the movie? Sure, fine, I guess I'll give it to you.

But it doesn't refute "they wanted to use a voice that sounded like the one in Her", and there are a number of indications the this was indeed the case.

> When the reaction is “I don’t care that an investigation refuted some of the accusations”, it demonstrates someone isn’t openly approaching things in good faith.

When the reaction is "it doesn't matter, it's still not ok to copy someone's voice and then market it as being that person's voice or related to that person's voice" and your reaction is to cast that as being something else, it demonstrates you are not openly approaching things in good faith.

An "investigation"?

Let's note that OpenAI didn't release the names of the voice talent since they said they wanted to protect their privacy...

So, how do you think the reporter managed to get not only the identity, but also the audition tape from "Sky"? Detective work?

An interesting twist here is that WashPo is owned by Bezos, who via Amazon are backing Anthropic. I wonder how pleased he is about this piece of "investigative reporting"?

> they said they wanted to protect their privacy

This very well could be a contractual obligation.

OpenAI allowed the reporter to hear some snippets from the audition tape. Not exactly my idea of an "investigation".

There are multiple parts to the voice performance of ChatGPT - the voice (vocal traits including baseline pronunciation) plus the dynamic manipulation of synthesized intonation/prosody for emotion/etc, plus the flirty persona (outside of vocal performance) they gave the assistant.

The fact that the baseline speaking voice of the audition tape matches baseline of ChatGPT-4o only shows that the underlying voice was (at least in part, maybe in whole) from the actress. However, the legal case is that OpenAI deliberately tried to copy SJ's "her" performance, and given her own close friends noting the similarity, they seem to have succeeded, regardless of how much of that is due to having chosen a baseline sound-alike (or not!) voice actress.

Have you listened to both voices in the comparisons floating around? There is no way any of SJ's closest friends or family would be fooled if that voice called them up pretending to be SJ.

Reply on news.ycombinator.com