Two subtitlers can translate the same line perfectly and only one of them is right. The difference isn't vocabulary. It's whether the viewer can physically read the thing before it disappears.
I spend more time fighting the clock than fighting the dictionary. A line comes in at 2.1 seconds. My Turkmen version is elegant, faithful, and 44 characters long. The reading-speed limit says 17 characters per second for adults, so that box holds about 35. Eleven characters have to die. Not "could be trimmed for flow" — have to die, or the subtitle turns red in the editor and the QA report bounces it back. That's the job nobody describes when they call subtitling "translation with timing."
The number that overrules your taste
CPS — characters per second — is the quiet dictator of the whole craft. Netflix caps adult content at 17, kids' at around 13. Some clients want 15. The Latin American Spanish desk I used to feed worked to 15 and it changed everything, because Spanish runs long and 15 is punishing. Turkmen has its own problem: agglutination. One noun collects case, possession, and plural suffixes and suddenly a three-letter English word is an eleven-letter tower. "To your houses" is one word, sure, but it's a long one. English gets to be terse for free. I don't.
So the real skill in subtitling isn't finding the right word. It's finding the word that means almost the same thing and fits. Those are different searches. The dictionary gives you the accurate one. The clock makes you take the shorter cousin — and knowing which cousins are safe to swap in, that's fifteen years of feel, not a glossary.
Here's the part that annoys me. Agencies review subtitles like they're reviewing a Word document. A reviewer marks my compressed line as "omission — source says X, target only implies X." Technically true. Also irrelevant, because the full X wouldn't fit in the box and the viewer would've missed the next line trying to finish reading this one. Omission is a feature of the format. If your reviewer flags every compression as an error, your reviewer has never subtitled anything.
Transcreation and subtitling pull in opposite directions
People lump these three things together — subtitling, transcreation, creative localization — like they're one skill with three hats. They fight each other.
Transcreation wants room. You get a marketing line, a pun, a rhythm, and the brief says "capture the feeling, forget the words." Great. I can rewrite the whole thing. I once spent an afternoon on a five-word slogan and delivered three options with back-translations and notes on why each lands differently in Turkmen. That's transcreation: few words, infinite time, paid by the hour or the concept because word count is a joke there.
Subtitling is the reverse. Infinite words coming at you, no room, and the clock — both the on-screen clock and the delivery deadline. When a joke shows up in a subtitle you don't get an afternoon and three options. You get 1.8 seconds of screen time and 32 characters, and the pun that worked in English relies on a homophone Turkmen doesn't have. So you don't transcreate the joke. You find something in the box that gives the viewer a reason to smile at roughly the right moment, and you let the clever original go. I've written about how the joke doesn't survive the character limit. It doesn't. Stop grieving it and hit the timecode.
The mistake I see agencies make is pricing and briefing subtitling as if it were transcreation. "Be creative with the humor!" Sure — inside 32 characters and 1.8 seconds, with lip-sync-adjacent timing constraints if it's for dubbing reference. Creativity with a gun to its head. That's a real skill and it should be paid like one, but it is not the open-field creativity of a slogan.
What I'd actually put in the brief
If you're a PM sending me subtitle work, three lines save us both a round of arguing.
Give me the CPS limit up front. Not "standard" — the number. 17, 15, whatever. It changes how I write from the first line, not after QA sends it back.
Tell me if these subtitles feed an AI dubbing pass or stay as reading text. If a voice engine is going to read them aloud, my constraints flip — timing to speech, not to reading speed, and suddenly the short compressed line sounds clipped and wrong. Different job. I need to know before, not after.
And tell your reviewer that compression is not omission. Put it in writing. Half the CPS-driven "errors" in QA reports are the reviewer scoring a subtitle by rules that belong to a legal contract.
The viewer never sees the timecodes, the CPS meter, the little red flags. They just feel it when a line stays too long or vanishes too fast. Get the number right and they notice nothing, which is exactly the point. Get it wrong and your beautiful faithful translation is the reason someone rewound.