How are you guys getting youtube training data at scale ?[D]

Working on a project where i need transcripts from a few thousand youtube videos as part of a training corpus. Mostly educational content, lectures, conference talks, that kind of thing.

Started with youtube-transcript-api (the python library) and it worked fine for the first couple hundred. Then youtube started returning empty responses. No error, no captcha, just a 200 with nothing in it. Rotated IPs through residential proxies and that bought me maybe another day before the same thing happened.

Tried yt-dlp for subtitle extraction next. Same story, works great until it doesnt, and when it breaks its always at 2am when your batch job is halfway done and you have no idea where it stopped.

The part thats frustrating is that these captions exist. Youtube has them. Theyre right there on the page. But pulling them programmatically at any real volume turns into this whole cat and mouse thing that has nothing to do with my project.

For the people who use youtube as a data source for training or eval, how are you handling this? Rolling your own infrastructure with proxy rotation? Using a paid service? Just accepting the failure rate and retrying? Or did you give up and switch to whisper on the raw audio?

Not looking for the theoretical answer, just want to know what works past a few hundred requests without babysitting it.