Most of us can’t read as quickly as we can absorb spoken language.

I learned this directly when I decided to create English subtitles for French films and TV as an exercise when I was trying to get from B2 to C1 in French. It was a good exercise, but the result was unusable because the text often goes by too quickly to read.

That’s when I understood that it really is an art.

The subtitle artist must make descriptions work, punch lines land, and reproduce dialog with the correct gravitas. And they have to do it while cutting 50 percent or more of a meaningful, culturally-grounded translation.

It doesn’t seem like the kind of thing an id-less ml model could ever do.