Ask HN: Why is uptalk intonation so prevalent in ChatGPT voices?
I’ve tried asking it to set voice with an even tone and less of the annoying uptalk but lately it just continues in this way. It hurts to listen to.
Plus I am sure that LLM engines could command a premium by shrewdly licensing such talent as James Earl Jones, Majel Barrett, or Milla Jovovich?
But it seems like the current trend is novel/generic voices in order to avoid suits/fees and pioneer new territory. Isn‘t Siri‘s personality a recognizable celebrity by now?
I reckon it took off because of telephones.
I always try to prefer virtual meetings with video and all the trimmings, especially screen-share capability. If not, in-person. Telephoning is like a last resort. I've pondered using TTY/TDD even though I have no physical impairments, but most agencies seem hostile to that sort of barrier and most of the time, I find myself needing to pass voice-recognizing gatekeeper AIs first.
When I was promoted to "sysadmin" in my first job they set me at a desk with a computer and a PBX telephone. (It was an office at a gov't defense contractor...) The computer I could relate to on an intimate level; the telephone scared the living bejesus out of me, and I prayed it never to ring.
A disembodied voice in your ear, especially one that interrupts everything you're doing and demanding a real-time conversation across space and time, that's incredibly rude, yet normalized for 100 years. I have PTSD and other disabilities, and I can get completely rattled and downright hostile/aggro when someone calls me and I'm not prepared, I'm not at my desk, no notes in front of me, even if I'm relaxing at home I just can't deal. I may answer the phone while on a bus or train, but I certainly won't sustain a conversation that way. Outgoing calls are sometimes OK but often end in disaster as I become increasingly frustrated.
Now it's not uncommon for someone who's standing at my front door to telephone me. What the hell, I'm less than 3 yards away in here, please don't send me across Cyberspace when you're manifesting in meatspace...
On a phone call I have no idea what office I've reached, what is a person's job title or role with a company, what their face looks like, when they may be taking a breath or using other physical cues to aid the conversation. So much metadata is lost in phone conversations that they are wholly dehumanized and require a revolution in etiquette, just to keep the peace.
Here's your uptalk origin: https://music.youtube.com/watch?v=Qb21lsCQ3EM&si=3j3bYd_0rPa... "Valley Girl, she's a Valley Girl!"
I'm not autistic and I can read expressions and body language, but perhaps autists are the evolutionary result of disembodied telephone conversations taking out the human element from verbal communication. It is simply amazing that even with apps and ubiquitous Internet and website portals that I still require the telephone to get important business done (and I'm always, always ticking "it's something else/other/Live Agent/Oh god this is an exception to all your rules please send me a human being").
It definitely makes reporting feel trustworthy and serious? When almost every statement sounds like a tentative question?
I've gotten old enough to now wonder if my dialect sounds like something from another world and era to younger folks in my region.
The way I felt about most of the Hollywood actors I heard from before technicolor was the norm.
I would be surprised if the majority of the training data is licensed from the speakers.
Often it's in a tonal particle, "One dim sum left meh." But it's possible in trying to artificially combine tone and text, the uptalk is moved up.
But the tell tones of a Malaysian accent is it's clipped. Instead of "I don't like that idea," it becomes "Don't like it." ChatGPT may be written American, so as an accent, it would sound closer to, "I- don't like, that idea."
And sentences often end in an elongated manner, "I wrote that is essay you wanted~". The elongated ends are quite common in many SEA accents as well, especially Thai.
Like ChatGPT's written english is or was close to nigerian business english...
I'm not aware of voice training though. x.AI outsources lots of stuff to Malaysia. Google has some, but has had this data like for TTS, STT, and Google Translate, for a decade or so.
/s
And so ChatGPT relies on the training data to know what that means so it leads to it talking like this as this is what the training data is filled with.
https://current.org/2015/06/a-top-audio-engineer-explains-np...
Is this ASMR?
Does it really? Isn’t this the kind of thing it would be good to practice tuning out rather than complaining about?
People who use this moronic phrase should go ahead and jump off a tall building...