A subtitle file landed yesterday with locked timecodes and a note: "Do not alter timing." The template was spotted for English. I was translating into Turkmen. Those two facts are at war, and the client didn't seem to know it.
This is the part of subtitling nobody warns you about. You think the hard problem is the language. It isn't. The hard problem is that someone already made the decisions that constrain your language, and they made them for a different one.
The pivot template is a house built for a different tenant
Here's how most of this work actually reaches me. A studio has a video. Somebody transcribes and times the English — the "template." That template becomes the source for everything: French, Portuguese, Arabic, and, once in a blue moon, Turkmen. The in and out times are fixed. The line breaks are suggested. The reading-speed cap sits there like a fence — usually 17 characters per second for adults, lower for kids' content.
And 17 CPS was tuned for English. Short words, tight morphology, a lot of meaning per syllable. Turkmen doesn't work like that. We're agglutinative. One English word becomes a stem with a suffix train hanging off it. "From our houses" is three words in English. In Turkmen it's often one long word, and that word eats characters. Case, possession, number, all stacked on the end.
So the cue that fit "I'll call you tomorrow" comfortably in English gives me a Turkmen line that blows past the CPS limit by a third. The QA tool flags it red. The client sees red and assumes I padded the translation. I didn't. The container is too small because it was measured for someone else's furniture.
What I actually do inside a locked cue
When I can't move the timing — and half the time I can't — the translation becomes an exercise in controlled loss. Not "what does this sentence mean" but "what survives in 1.8 seconds of screen time for a Turkmen reader who's also watching a face and a plate of food."
A few things I reach for, in rough order:
- Cut the vocatives and the discourse fillers first. "You know," "well," "look" — English scripts are full of them and Turkmen carries the same weight through word order and particles. Drop them and nobody misses anything.
- Trade a suffix chain for a shorter construction when one exists. Sometimes a postposition reads faster than a fully inflected case ending. Sometimes not. You learn which by counting, not by feeling.
- Let the picture do work. If the character is clearly handing over money, I don't need the verb spelled out with all its politeness marking.
What I refuse to do is fake compression by machine-flattening the register. Turkmen has real distance between formal and casual address, and English templates are blind to it. If I collapse everything to neutral just to hit CPS, the subtitle reads correct and feels dead. That's not a saving. That's the transcreation quietly dying to make a number go green.
The timing note that should be a negotiation
"Do not alter timing" is the line I push back on the most. Sometimes it's non-negotiable — burned-in captions, synced to a hard cut, forced narrative. Fine. But a lot of the time the studio wrote that note because moving cues is annoying for their conform process, not because the physics demand it. Nudging an out-time by 300 milliseconds into a pause costs them nothing and gives me the breathing room to say the sentence properly.
So I ask. Cue by cue where it matters, with the timecode listed. "Cue 214, requesting +0.4s to out — Turkmen construction runs 22 CPS otherwise, here's the alternative if timing is truly frozen." Give them the fix and the fallback in the same breath. Nine times out of ten they say yes, because you've made it easy and you've shown you actually tried to fit it first.
The automatic-subtitling crowd sells this as solved — cheap AI drafts, a human tidying it up fast. And for the English template? Sure, machine spotting plus a cleanup pass is genuinely useful now. But that entire pipeline is measured against Latin-script, low-morphology languages. The tools don't know that a language exists where the reading-speed math breaks. They'll produce a Turkmen line that scans as fine to the algorithm and rushes the viewer off a cliff. The "human tidy" they budget an hour for is where all the real subtitling actually lives.
If you're a PM sending me a subtitle job, the single most useful thing you can tell me isn't the CPS cap. It's whether the timing is genuinely locked or just inconvenient to change. One of those I work around. The other one I fix — if you let me.