How the voices for ChatGPT were chosen

[datahack]

Am I the only one that really dislikes these overly perky voices? And the one they used for the demo was GOD AWFUL. If this is what the future sounds like… idk. I was hoping for a calm, neutral and not eager assistant with a little class. A little Jarvis maybe? Something calm and easy on the ears.

These all sounded to me like people auditioning for a role in film and television or something.

[DeathArrow]

I would like something more David Attenborough-ish.

[vlasky]

I'm surprised that you can't configure the pace or pitch of the voices. The Web Speech API, around for a long time, has that capability. I'd also love to have the option of Australian and British accents.

[iLoveOncall]

I use "Cove" which is much more tame that the other by default and (I assume that's the case for all) matches your tone.

So if I'm asking it a question on a neutral tone it will answer in the same way.

[exitb]

This is contextual, the model seems to match the user vibe and the general situation. Take a look at the "Customer service proof of concept" video added to the announcement post[1] - there are two AIs talking using very different voices.

[1] https://openai.com/index/hello-gpt-4o/

[genericacct]

You're not alone. But it should be an easy fix with voice cloning proceeding as fast as it is.

[jimsimmons]

Google did a much better job here and everyone glossed over

[jareklupinski]

i want jonah jameson from the spiderman movies yelling things at me

[LeoPanthera]

None of the voices really seem that good to me. I think it's because they're all aggressively American. It would be nice to hear some other accents.

[Joeri]

I had Sky answer in dutch, and hearing dutch with an american accent is funny but a little jarring. The frustrating part of being a flemish dutch speaker is that dutch itself is already poorly served by AI tools (it’s actually impressive chatgpt can do it at all) and additionally the flemish accent and dialects are not supported at all in pretty much anything, so all dutch text to speech sounds foreign.

I hope eventually they can pick up language, dialect and accent from youtube content or something like that, because voice casting will never scale to the thousands of variations they need to have to support a global audience. Instead of picking a voice from a handful of choices it should do a q&a where it learns your locale and intonation preferences and generates a tailored voice on the fly.

Still, I am impressed with the quality of the voices they have and with how well they manage dutch. It is just the case that there still is a lot of room for growth.

[nacnud]

As a British person, I find all the current OpenAI voices have overly strong (American) accents, which are way too perky/enthusiastic for my ears. It would be great if they could offer a more neutral accent, or even one British accent? (Example: Voice 5 from pi.ai is excellent, IMHO)

[glimshe]

Some British accents, in particular outside of London, can be almost unintelligible to general English speakers.

There is nothing intrinsically neutral about the British accent, it's more a matter of diction from the speaker. I would also not consider the OpenAI accents as neutral American, feels more like "Young Californian adult woman" accent to me.

[msoad]

Berlin famously casted a gender neutral voice for public transport announcements. It would be nice if OpenAI would have done that too. At least for one of the options. Computers better not have genders imo

[Spiwux]

They're clearly trying to imitate human interaction, and believe it or not, the vast majority of humans have genders they're perfectly happy with.

[arianvanp]

https://www.tiktok.com/@openlynews/video/7260846856228326682...

If you're interested in her voice

[xdennis]

> Berlin famously casted a gender neutral voice for public transport announcements.

The person in question is transsexual. You're saying that regular men and women aren't neutral but transsexuals are?

> Computers better not have genders imo

This doesn't make any sense. Gender is an intrinsic part of many languages which cannot be ripped out. It's impossible to talk in some languages without picking a way of speaking which is either masculine or feminine.

[avereveard]

I'd love a glados voice or a extremely synthesized one 80s style. I'm using an open one for now, and I think it's very appropriate.

[AnonymousPlanet]

I wouldn't be surprised if the person voicing those announcements would be more happy to be called female than gender neutral.

[GaggiX]

What is a gender neutral voice? Something that is interpolated between a male and a female voice?

[iLoveOncall]

Why?

[iamflimflam1]

Article doesn’t address the ridiculously giggly female voice they used for the 4o demos.

[isoprophlex]

Fake, americanized, subtly sexualized. I'm pretty sad that they went for the lowest common denominator approach.

I'm not opposed to computers using emotion at all, mind you. But I don't like that arguably the AI company furthest ahead is choosing this gaudy hollywood approach to marketing.

[riffraff]

Desi Lydic on the Daily Show had a perfect take on this

https://www.youtube.com/shorts/51ucQ4s7Crc

[thih9]

> We support the creative community

I doubt the net result is support. Expanding creative community - yes. But support - not quite; established creators will likely suffer because of openai, most of them already having a harder time.

[gaymenexisttoo]

Sky has an uncomfortably flirty tone to me. If they're going to go that route, at least include an uncomfortably flirty male voice too. (Or admit Sky is there more for the enjoyment of straight male users than anything else)

[saurik]

> We believe that AI voices should not deliberately mimic a celebrity's distinctive voice—Sky’s voice is not an imitation of Scarlett Johansson but belongs to a different professional actress using her own natural speaking voice. To protect their privacy, we cannot share the names of our voice talents.

I love how they have to go out of their way to explain that, rather than train a model on Scarlett Johansson's voice, they came about her voice honestly, by just finding someone who's natural voice sounded the most like Scarlett Johansson... as, clearly, no one is questioning that this voice was chosen to mimic Scarlett Johansson.

[Havoc]

Pretty sure that’s another joking reference to the movie Her rather than actual concern about similarity. If that was the case pointing it out would attract lawsuits

[defrost]

Some are questioning why they chose not to mimic Stephen Fry https://www.youtube.com/watch?v=J7E-aoXLZGY or Attenborough.

[th0ma5]

From their Discord server: > @ everyone We’ve heard questions about how we chose the voices in ChatGPT, especially Sky. We are working to pause the use of Sky while we address them. Read more about how we chose these voices [link above, but no further details]

[easymodex]

Wow, so they actually got bullied into removing the voice? What a world. I actually liked that voice or at least didn't find any issue with it, they are all perky assistant type voices, but why not? Adding extra choices is fine if people want it, but removing the voice due to loud minority effect is just sad.

[omarfarooq]

Would it be possible to use prompting to change aspects of the voice output? Will the voice respond angrily if you ask to act an angry character in a play?

[prolyxis]

Yes. This was a big part of the GPT-4o demos. But likely they've specifically fine-tune it not to show negative emotions in a serious fashion.

[iLoveOncall]

You don't even have to prompt it, it will match your tone.

[input_sh]

I can't help but read this as:

> 400 people did some unpaid work for us, only 5 of which eventually got paid, after 5 months of not knowing will they or won't they.

Fair enough, that's how things work for other voice acting jobs as well, I just doubt it usually takes almost half a year. I wonder how many of the shortlisted 14 have just given up somewhere along the way.