I once caught a subtitle file that rendered a five-second line of Korean dialogue as 'The situation has become rather more complicated than we had initially anticipated.' The actor was screaming. The scene was a car chase.
That's what bad subtitling looks like. Not mistranslation. Not grammar errors. It's what happens when a subtitler is working from a script with no video reference, against a deadline that was already impossible when they accepted the job, and nobody told them the show was an action thriller.
I need to explain something about how subtitling actually works on streaming platforms, because most of what people think they know about it is wrong. The job is not 'translate the dialogue.' The job is: watch the show, understand what each line is doing narratively and emotionally, translate it to preserve that narrative and emotional function, and compress the result into a string of text that a viewer can read in the exact number of seconds the original actor took to speak the line. Do this roughly 800-1,200 times per episode, depending on the script density. Do it fast enough that the episode can release globally on the same day it drops in Korea or Japan or Spain. Do it without ever being noticed, because the moment a viewer notices a subtitle, you've already failed.
And right now, the entire industry is set up in a way that makes doing this well nearly impossible at the speed and volume the platforms are demanding.
Let me tell you about a job I watched go wrong.
Korean thriller series. The kind that everyone's trying to make since Squid Game blew up — dark, tense, morally complicated. Actually good. Streaming platform picked it up for international distribution and decided on a global day-and-date release. That meant English subtitles had to be locked before the Korean domestic premiere.
The subtitling house got the files two and a half weeks before the deadline. Not unreasonable for a 12-episode season if you have a team of five or six subtitlers working in parallel. They had three.
Two of the three subtitlers didn't get video access until week two because the post-production team was still locking the final cut. The subtitling house's project manager apparently decided — I don't know if this was their call or the platform's, I never got a straight answer — that the subtitlers could start from the dialogue transcripts alone and time the subtitles later when the video was available.
If you've subtitled professionally, you already know where this is going. If you haven't: translating from a transcript without video means you don't know who's speaking to whom. You don't know if the line is delivered in a whisper or a shout. You don't know if it's playing over a quiet conversation or a gunfight. You don't know where the shot boundaries are. You're translating blind.
The subtitles went live on schedule. The show was good — genuinely good. But the English version had this weird, slightly off quality. Dialogue that was clearly intense in Korean was rendered in the same flat register as exposition. Characters who were supposed to sound completely different from each other — a grizzled detective, a young prosecutor, a crime boss in his 60s — all used the same vocabulary and the same sentence structures. It wasn't wrong. It was just... same-y. Like someone had run the whole script through a language normalizer.
Twitter noticed within hours. The show's international reception was fine — it wasn't a disaster. But it could have been a phenomenon. The source material was strong enough to carry it. The subtitles didn't ruin it. They just took everything that was distinctive about the show's dialogue and smoothed it into something generic.
That's the thing people miss about bad subtitling. It's rarely catastrophically wrong. It's rarely Google Translate gibberish. It's just... personality-free. And in a medium where dialogue is the primary vehicle for character development, personality-free subtitles are a form of content degradation.
The CPS constraint that governs everything
Here's the technical reality that nobody outside the industry thinks about, and it controls every subtitle decision.
A human being can read subtitles at roughly 17-20 characters per second. That's for Latin-script languages. Different scripts have different limits — CJK languages (Chinese, Japanese, Korean) are measured differently because a single character can represent an entire word, so the metric shifts to characters per line rather than per second. But for English subtitles, 17 CPS is the standard ceiling. Some platforms push to 20. Netflix uses a slightly different system based on words per minute, but it maps roughly to the same territory.
A single subtitle line maxes out around 37-42 characters. Two-line subtitles are standard. Three-line subtitles are technically possible but strongly discouraged because they eat too much screen real estate and read slowly.
So the math is: if an actor delivers a line in 2.8 seconds, and your CPS ceiling is 17, you get roughly 47 characters including spaces to render that line in English. That's about seven or eight words. Seven or eight words to convey everything the original dialogue conveyed in however many syllables the actor used in Korean or Japanese or Turkish.
Now imagine a line where the original is compact and the English equivalent is inevitably longer. Japanese-to-English is especially brutal for this because Japanese can pack multiple layers of meaning — subject, object, verb, honorific register, speaker attitude — into a structure that's much shorter than its English equivalent. A seven-syllable Japanese line might need 12 words in English. Your 47-character window just got impossible.
Now imagine this happening 800 times per episode. Now imagine needing to do it not just accurately but with character differentiation, tonal consistency, cultural adaptation, and shot-change-aware timing.
This is what people are actually paying for when they pay for subtitle translation. Not vocabulary knowledge. The compression decisions. The hundreds of micro-decisions per episode about what to keep and what to sacrifice.
What compression actually looks like
I want to walk through a real example because otherwise this sounds abstract, and it's the opposite of abstract. It's the most concrete, granular, word-by-word decision-making in any translation discipline I know of.
Korean thriller. Scene: a warehouse confrontation. The protagonist — a detective who's been investigating a corruption case — has just been ambushed. He's injured. His former partner shows up, revealed to be working for the people the detective is investigating. The detective says one line. It's loaded. It has to carry betrayal, anger, exhaustion, and the specific history of their relationship.
Literal translation of the Korean: 'After everything we've been through together over all these years, how could you look me in the eye and do this to me right now at this moment?'
Character count: 140.
Screen time: 3.1 seconds.
Available characters at 17 CPS: ~52.
So you have to collapse 140 characters of literal meaning into 52 characters of subtitle text while preserving: the betrayal, the history between the two characters, the power dynamic (the speaker is injured, the person he's addressing has the upper hand), and the emotional register (not just sad — angry, exhausted, disbelieving).
A bad condensation would be something like: 'After everything, how could you do this?' That's 42 characters. It fits. It's grammatically correct. It's also completely generic. Replace 'detective betrayed by his partner' with 'person disappointed by their roommate eating their leftovers' and the line still works. That's how you know it's wrong. A good subtitle line is specific to the scene it's in.
A better condensation: 'All those years. And you look me in the eye.'
That's 47 characters including spaces and punctuation. It drops the explicit question ('how could you') and the explicit betrayal framing ('do this to me'). But it preserves the history ('all those years'), the confrontational intimacy ('look me in the eye'), and the gut-level sense of violation. More importantly, it preserves the narrative function: the detective, injured, on the ground, refuses to break. He answers betrayal with confrontation.
Could you do better with more time? Absolutely. Give me a full revision pass and I could probably shave the line differently and preserve slightly more of the original. But this is the constraint: the condensation decision is being made for every single subtitle in the episode, and the subtitler is making each of those decisions in maybe 30-60 seconds because the episode has to ship.
The best subtitlers I've worked with don't think about this as translation. They think about it as narrative compression. The question isn't 'what does this line say.' The question is 'what does this line do in the scene, and how do I make the English subtitle do the same thing with half the words.'
The platform QA blind spot
There's something wrong with how streaming platforms evaluate subtitle quality, and it's been bugging me for years.
The standard subtitle QA checks are all technical. Is the timing accurate? Is the CPS under the threshold? Are there any formatting errors in the file? Any overlapping subtitles? Any subtitles that violate the minimum duration? These are all automatable and they're generally well-automated at this point. A subtitle file that passes these checks is technically compliant.
The problem is that technical compliance says nothing about whether the subtitles are any good. A subtitle file can be perfectly timed, perfectly formatted, CPS within limits, all technical boxes checked — and still make every character sound like a customer service chatbot because the translator wasn't differentiating voices.
I once reviewed a Spanish-language drama's English subtitles where a dozen different Spanish idioms had all been rendered as some version of 'unbelievable.' A grandmother in her 70s: 'Unbelievable.' A teenage boy: 'Unbelievable.' A corporate lawyer: 'Unbelievable.' The file passed QA. Technical compliance: perfect. Viewers noticed immediately because viewers aren't stupid. The show got dunked on.
Linguistic QA — actually watching the content with the subtitles and evaluating whether they work — is expensive. It requires native-level reviewers in both languages who understand the content genre and have enough time to actually watch the show. For a 12-episode season in 30 languages, that's 360 person-hours of review time, minimum, if each language gets one reviewer doing one watchthrough. Most platforms aren't spending that. So linguistic QA gets compressed into spot-checking, and spot-checking catches formatting errors but doesn't catch the thing where a detective, a prosecutor, and a crime boss all sound like they were written by the same intern.
This isn't a problem the platforms don't know about. They know. The tension is that linguistic QA at scale costs real money and takes real time, and the two structural demands of streaming subtitling — more languages, faster turnaround — make linguistic QA more necessary and harder to afford simultaneously. I don't know what the solution is for the industry as a whole. I just know that as an individual subtitler, the only defense is to give the translators what they need to get it right on the first pass, because the second pass might not exist.
Stuff that should be standard and isn't
I'm not going to give you a checklist. I hate checklists. They make people think the work is simpler than it is. But there are some things that make the difference between subtitles that work and subtitles that don't, and they're all process decisions, not translation decisions.
The subtitler needs the video. Not the script. Not the dialogue list. The actual show. You cannot subtitle something you haven't watched because you don't know where the shots cut, what the actors are doing with their faces, whether the line is whispered or shouted or sobbed. I've had jobs where the client sent a transcript and said 'the video is still being edited, start now and we'll adjust later.' That's like asking someone to cook a meal without telling them what the ingredients are. Every time someone agrees to subtitle from a transcript, the output gets worse. It's not a skill issue. It's an information issue.
Different genres need different subtitle strategies. Comedy timing is completely different from thriller timing. In comedy, the subtitle has to land on the punchline beat. If it arrives early or late, the joke dies. In an action scene, subtitles should be shorter and faster — the viewer's attention is split between the text and the action, so you need even more aggressive compression than usual. In a period drama, you might need to preserve a slightly more formal register. The subtitle conventions for a K-drama romantic comedy are not the same as the conventions for a Scandinavian noir thriller. A style guide per genre per language pair sounds like a lot of documentation. It's less documentation than the time you'll spend fixing subtitles that used the wrong conventions.
Season-long glossaries. A 16-episode K-drama. Episodes 1-4 done by Subtitler A. Episodes 5-8 by Subtitler B. Episodes 9-12 by Subtitler C. Episodes 13-16 by Subtitler D. If they're not working from the same English equivalents for recurring Korean terms, the subtitles drift. The same honorific gets translated three different ways over the course of the season. The same workplace jargon switches between English equivalents. Viewers definitely notice. The fix is a shared glossary distributed at the start of the project and updated as the season progresses. It's not complicated. It's just not always done.
And the timeline is not a variable you can optimize your way out of. Subtitling takes the time it takes. You can compress it to a point, and beyond that point, quality collapses. I can't tell you exactly where that point is because it depends on the content and the language pair, but I can tell you that if someone is offering you 12 hours of content subtitled into English in five working days, the subtitles will not be good. They might technically work. They won't be good. And the audience will know.
The platforms that take subtitling seriously — and some of them do, to be fair — build these things into the workflow from the start. Template files with locked timing. Per-language style guides by genre. Shared glossaries. Video access for every linguist. Realistic timelines. It's not rocket science. It's just process design. The difference is that the platforms that do this see subtitling as part of the content production pipeline, not a post-production afterthought to be squeezed into whatever time is left before the global release date.
Artlangs Translation does OTT and VOD subtitling in 230+ language pairs. Template-based workflows with locked timing, per-language and genre style guides, season-level glossaries, dual-pass technical and linguistic QA, video access for every subtitler. If your subtitle problems are process problems disguised as translation problems, the fix is in the workflow, not the word choice.
