Of late I’ve been listening to the “Sleepy Science” channel on Youtube. There are a zillion similar channels with similar names; the one I’m referring to produced this video, among others..
I like the channel, in that it seems to be fairly accurate (I haven’t found any glaring inaccuracies, to the extent that I’m knowledgeable about any of the topics covered), though I have heard some words badly mispronounced, unless the British pronunciation would be very different than what I’m used to as an American. The material that is read is soothing and quite repetitive; perfect for absorbing new information if you manage to stay awake, but better yet, it’s easy to fall asleep to.
Question: is this AI generated? I’m pretty sure the voice is, as the volume of material seems unrealistic for a single live human to create. The intonation is a little “off” sometimes, too, although for an AI voice it’s pretty good.
And, where does the content come from? From sentence to sentence, it tends to sound pretty natural, but the organization is strange; sometimes it sounds like the voice is reading a subheading (without realizing it’s a heading and making appropriate tonal adjustments). And the content will meander and, using different sentences, repeat the same facts.
So - is this all AI? An if it is, should I care? It’s wonderfully soporific, and I can easily create 8+ hour playlists to drone all night while I sleep much more calmly than I otherwise would.
Sounds fairly natural to me (a random British bloke). Could you give some
examples of the mispronunciation (probably just a British/American thing) and
inappropriate tonal adjustments - I had a quick listen to some of it and it
sounds very good if it is AI !
I am quite sure that all the videos on that channel, and many other similar channels, are entirely AI-generated. I had some questions about the channel over here, but no one has answered them yet.
The first few nights [my wife] played them, I marveled at the notion that someone had written and narrated (very professionally) such long videos. It was only a few nights later, when it mispronounced a word (I think it was “live” with a short I instead of a long one, or some such thing) that a real human would never do, that I finally realized, “Oh, this is all AI.”
FWIW, in that portion of the videos I have listened (and paid attention) to, I have not heard anything obviously wrong within the limits of my understanding of the various topics.
Since then I’ve found that there are lots of other similar channels of entirely AI-generated videos for sleeping. To say nothing of all the awful AI slop of other kinds, clogging up YT.
(The putative creator of the site explains how he came to do it in this video, but a comment suggests that that video is also AI! It’s AI all the way down!)
Oh wow, I hadn’t seen your post in the Pit. We certainly have the same take on that channel! The video you link to doesn’t help much, does it? I tend to think the guy is real but the “help” he vaguely refers to is AI.
I will try to jot down some examples as I come across them and post them in this thread. They are fairly infrequent, and I don’t specifically remember any off the top of my head (not surprising, since my goal while listening is to fall asleep rather than think too hard).
I guess my main question is, if it is largely AI (and it’s got to be, no human could produce so much material on so many subjects so quickly without help), is it harmful in any way or is it okay? If the channel “creator” had instead invented a new kind of white noise machine and was selling those to help people sleep, I wouldn’t find that immoral. Is this any different?
ETA: I just read some of the comments on the “why I made this channel” video. There seems to be consensus that it is indeed narrated by AI. I guess it’s obvious that it is; not sure why I had any doubt.
Two examples that I have noticed are the words record and present, which have slightly different pronunciations depending on the meaning. When you record audio, for example, it is pronounced with a long e. Back in the day, if you wanted to buy the latest album from your favourite band you’d go to a record store, pronounced with a short e. If I give someone a present the first e is short, but if I present them with an award, its a long e.
In England, it’s slightly different…
The nouns record and present would have a slight emphasis on the first syllable.
The verbs record and present would have a slight emphasis on the second syllable.
Neither would have a long e sound.
I think both sets of pronunciations - @moes_lotion 's and @pjd 's - work in American English. What pjd describes is closest to how I would say the words, but moes_lotion’s version also sounds normal to me.
There are many words with the pattern that they are accented on the first syllable when used as a noun but on the second (or a later) syllable when used as a verb: record, present, project, permit, finance, suspect, conflict, combine, insult, implant, reject, relapse, produce, transform, escort, increase, etc. This process is called “initial-stress derivation”. Wikipedia says there are more than 170 such pairs.
I can imagine that an AI reader might not get the pronunciation right with such words. However it doesn’t seem like a huge difficulty for an AI to determine whether a word is used as a noun or verb in a sentence, so I would guess that the ever advancing sophistication of our electronic overlords will soon make this a problem of the past.
I haven’t noticed a “reCORD/REcord” type issue in the YouTube channel I mentioned in the OP. It was more likely to be unusual words - I’ll list some if I notice in the next few days.
One dead giveaway to AI narration, in addition to the mangled names and flubbed pronunciations, is that the narrator can’t deal with numbers at all. So, say, 1200.00 will be read as one-two-zero-zero-period-zero-zero.
Interesting! I have not heard that problem, but I am guessing he prepares the narrative so that this doesn’t come up.
I don’t know how he puts the material together, but it is quite poetic in spots and sometimes has first-person observations (not often). I am guessing he feeds a ton of material to an AI processor of some sort and tells it, “produce a narrative about [tsunamis, butterflies, chocolate, whatever] that is about two hours in length if read out loud. Spell out all numbers.”
Probably other instructions as well. Avoiding repetition is not one of them, though. Last night I listened to the cave episode (or part of it - I did fall asleep) and I think there are at least three mentions of some stalactite that is 200 feet tall or something.
Not all YouTube channels are created equal - it’s possible to collaborate with AI and produce a good result, as long as there is a human curating the content. I think that’s more or less what channel I referenced does.
I am certainly not an expert on all the topics I’ve listened to, but to the extent that I do know stuff, it seems that this channel uses reputable sources for information gathering.
Not that I’m defending the practice in general, but I will note that a knee-jerk “arghh, AI was used, it must be ‘nonsense’” reaction is unnecessary.