Hacker News Reader

On a tangential note: As far as training data for text and speech driven machine learning models go, isn't srt files truly remarkable?

They are time stamped hence annotated precisely, hash and file matching makes media content to subtitle matching exact and overall there are language variations.

View on HN

They are very artisanal, though. I used to do timing for anime fansubs when I was a teenager, and there's a lot of subtlety going on, specially when you don't speak the language (you're guided mostly by sound and phonetics, as the translations come more or less untimed).

Timing is much more automatized nowadays, but still doesn't feel as high quality as handicrafted timing does.

Reply on news.ycombinator.com