What is genuinely going on with YouTube subtitle transcriptions?
I mean aren't they like one of the top 3 AI companies in the world? I didn't even think it was allowed to say that word. It bleeps "fuck" and "shit", like...
Hey, Grok, is the expression "to get raped over the coals"?
5 replies
This is NOT a one off incident, anyone who uses subtitles, correct me if I'm wrong. Like, it goes out of its way to make the most incomprehensible interpretation of anything that is said, it's not even good at transcribing what is said, it just actively throws the most random shit it could in there.
I have had text to speech on my phone for a decade and it has never sucked as bad as YouTube's fucking captions.
It's voice to text. It's only off by 1 letter. Way worse with people who have accents.
Watched an amateur boxing tournament on there not too long ago. It turned this guy's name, Orlando Holley-Sotomi into "Holy Sodomy." https://boxrec.com/en/box-am/1075223
But I mean, even their own filters don't allow the word "rape", and every single other STT has worked fucking great (-ish) forever, but this is just horrific. The vanilla Android STT works well enough... Sure it does it in realtime, but, if that is the case- WHY? Surely there is a more programatically expedient way of solving the problem? Like, the first query for the STT could be stored for that video, so that the next user who comes along and wants subtitles get served the same text, and then meanwhile, just run that through their own fucking AI to make sure it makes sense and is coherent, and exchange the previous subtitles with the new ones?
Why doing all this impromptu on the fly shit per instance at all?
Is this what they need all their fucking data centers for??
Lack of cache???
Running every YouTube video through AI just to transcribe it would take a huge amount of resources. That's probably the only reason. Whatever the solution, it needs to be something that can be done quickly across a billion videos.
My point is, you only need to run it once. Instead of every time the video is played.
Surely the former is better.
Just run the audio through the STT, cache that data. Run it through AI once for coherence. Store in cache.
Every time anyone requests the video subtitles, retrieve cache.
Am I insane?
Edit: Thanks for interacting though, I'm not mad at you, I'm mad at the world, and it shall pay, pay dearly, one day. One day they'll see. One day they'll not be laughing anymore! Just you wait. JUST YOU ALL WAIT.
Fuck greenshot for capturing the mouse. I know I can fix that in preferences but I am too lazy, it should be perfect from the start