A pun landed on my screen last week: a two-word English gag built on a homophone, timed to a facial reaction that lasted maybe a second and a half. The subtitle window gave me 38 characters across two lines and a reading speed I couldn't blow past. Turkmen has no matching homophone. It never will. So the real question wasn't "how do I translate this" — it was "which thing do I sacrifice, and can I hide the seam."
That's the whole job in creative localization, most days. Not finding the perfect equivalent. Deciding what to lose.
Transcreation wants room. Subtitling takes it away.
These two disciplines pull in opposite directions and nobody in the brief ever admits it.
Transcreation is the luxury mode. You get a source line, a note about tone and intent, and permission to walk away from the literal words entirely as long as the effect survives. Marketing pays for that. It assumes space — a headline can breathe, a tagline can be reworked five times until the client's gut says yes. When I transcreate a slogan into Turkmen I'm building for feel: rhythm, the sound of the words out loud, whether it reads like a human wrote it or a form field generated it.
Subtitling is the opposite. It's transcreation wearing a straitjacket. Same demand — preserve the effect, the humor, the register — but now I've got 42 characters per line, two lines max, and a minimum on-screen duration a Turkmen reader needs to actually process the words. Turkmen runs long. Agglutination stacks suffixes onto stems until a compact English idea becomes a caravan. "You should have told me" is four short English words; the Turkmen carries the mood, the tense, the person, the reproach all bolted onto one verb, and it doesn't shrink to fit.
So when a transcreation instinct meets a subtitle box, the box wins. Every time. The skill isn't fighting that — it's choosing your loss on purpose instead of letting the character counter choose for you.
What I actually cut, in order
I've got a rough hierarchy I run in my head. It's not sacred but it's saved me a lot of second-guessing.
First to go: the mechanism, keep the payoff. If a joke works because of a specific wordplay Turkmen can't reproduce, I stop trying to be clever with the sounds and go for the reaction the joke is supposed to produce. Make the viewer feel the beat lands, even if the gears underneath are completely different. Nobody in Ashgabat is auditing whether my Turkmen pun uses the same phonetic trick as the English. They're checking whether they smiled at the right moment.
Second: register before content. In a heated exchange, how someone says something often matters more than the literal words. A character being coldly polite while furious — that tone survives even if I compress the actual sentence to half its detail. Turkmen has real tools for this: the respectful versus blunt forms, the softening particles you drop to sound curt. I'll spend my scarce characters on getting the temperature right and let some of the informational content evaporate. Context and picture cover the gap.
Last resort: split the idea across the cut. If a line genuinely won't compress, I look at whether the previous or next subtitle has slack. Sometimes you can plant setup one card early so the payoff card only carries the punch. This works maybe a third of the time and requires you to actually watch the timing, not just translate the SRT rows in a spreadsheet like it's a glossary.
The thing I refuse to do: cram. A subtitle that technically fits the character count but flashes past faster than anyone can read is worse than a slightly looser translation the viewer can actually finish. A subtitle nobody reads in time isn't a subtitle. It's decoration.
The AI dubbing question nobody asked me
Here's where this gets pointed. Auto-dubbing and machine subtitling are being sold hard right now — clone the voice, generate the timing, done. And for informational content, straight expository stuff, factual narration, I'll grant it's coming for a chunk of the work.
But the machine doesn't choose a loss. It doesn't know the pun is a pun. It renders the literal meaning, fits it to the timecode by brute force, and produces Turkmen that's grammatically fine and completely dead. Fluent and wrong, the same failure I keep running into. It'll translate the mechanism and drop the payoff without knowing there was a payoff. It has no hierarchy. It just fills the box.
And Turkmen was never dubbed at scale to begin with — there's no decades-deep tradition of localized voice work here for a model to have learned from. So the training data for "how does a Turkmen villain sound menacing but restrained" mostly doesn't exist. The machine will guess, confidently, and the guess will be a flat reading of the words.
That gap — the deliberate, informed sacrifice — is the part of this work that's still mine. Not because I'm precious about it. Because the tool literally cannot see the choice that needs making. It sees 38 characters and a timecode. It doesn't see the second-and-a-half of a face waiting for a laugh.