well, i have no evidence of this. however. looking at the way auto-generated subtitles are served at youtube right now, they are sent individually word-by-word from the server, pick up filler words like “uh”, and sometimes pause for several seconds in the middle of sentences. and they’re not sent by websocket, which means they go through multiple requests over the course of a video. more requests means the server works harder because it can’t just stream the text like it does the video, and the only reason they’d do that other than incompetence (which would surely have been corrected by now, it’s been like this for years) is if the web backend has to wait for the next word to be generated.
i would love to actually know what’s going on if anyone has any insight.
Andrew Ng did a video when he gradually added noise to the training audio to improve the quality.
But here we are dealing with homophones so it’s not just turning speech to text, it also needs to be context aware.
Possible but too expensive to implement automatically.
context awareness is the entire point of language models tho :(
I’m highlighting that speech to text and context awareness are different skills.
YouTube is unlikely to waste loads of compute power on subtitles that don’t need it just to capture the occasional edge case.
i mean, it’s a one-time-per-video thing. they already do tons of processing on every upload.
So if you can reduce compute there then you save money.
There is no technical difficulty. It’s a business decision.
right now they’re dynamically generating subtitles every time. that’s way more compute.
For real? That’s incredibly dumb/expensive compared to one subtitle roll. Can you share where you saw that?
well, i have no evidence of this. however. looking at the way auto-generated subtitles are served at youtube right now, they are sent individually word-by-word from the server, pick up filler words like “uh”, and sometimes pause for several seconds in the middle of sentences. and they’re not sent by websocket, which means they go through multiple requests over the course of a video. more requests means the server works harder because it can’t just stream the text like it does the video, and the only reason they’d do that other than incompetence (which would surely have been corrected by now, it’s been like this for years) is if the web backend has to wait for the next word to be generated.
i would love to actually know what’s going on if anyone has any insight.
it would be an improvement. thats not what we are doing anymore
new tech is there to make everyone more miserable