Subtitling a ten-minute video used to mean a transcriptionist, a wait of days, and a translator per language, at a price that ruled it out for anyone without a budget line for it. In 2026 the whole chain runs in minutes on a laptop. That changes the question. It is no longer whether you can afford subtitles in ten languages; it is whether anyone will check them, because the models now produce output that looks finished and is not.
I do not subtitle video for a living, so take the workflow and not the creative advice. What I do know is deliveries. A subtitle file is a data file with a format, timings and a reading-speed budget, and it can be checked against rules instead of by watching the video and hoping. That is the angle of this post: the pipeline end to end, the formats, where translation breaks, and the handful of checks I would run on every file before it goes up.
Three models, three places it goes wrong
The pipeline has three stages, and each is a different model with a different failure.
Speech recognition turns the audio into text with timestamps. OpenAI's Whisper (free, runs locally), Google's Chirp and Deepgram are the usual choices. On clean audio from one speaker the mistakes are a few per hundred words; on two people talking over each other, an accent the model saw little of, or a room with a fan, they multiply. Names and product terms are the reliable casualties.
Segmentation cuts the transcript into subtitle blocks: one or two lines, roughly forty-two characters per line, on screen for as long as a reader needs and gone shortly after the speaker stops. A block should not split mid-phrase, and two speakers should never share one. This is where tools differ most, and it is invisible in a transcript review because the words are all correct.
Translation then rewrites each block into the target language while keeping its timing. A six-word English block can become a twelve-word German one, and it still has the same two seconds. The timing was set for the source language, and nothing in the pipeline resets it for the target.
The rule I check first is reading speed: about twenty characters per second for adult viewers, which is the figure the streaming platforms work to. Paste a translated block into the Word Counter for its character count, divide by its duration, and any block over twenty is one the viewer will not finish.
SRT for uploads, VTT for your own site
The content of a subtitle file is the same everywhere; the framing differs by format.
SRT (SubRip) is the plain one: a number, a time range, the text, a blank line. Every platform accepts it and any text editor can fix it.
`
1
00:00:01,000 --> 00:00:04,500
Welcome to this tutorial on
building your first API.
2
00:00:05,000 --> 00:00:08,200
We will start with the basics
and work our way up.
`
WebVTT is the web version, required by the HTML5 element, and adds positioning and styling. Note the full stop in the milliseconds where SRT has a comma; that one character is the most common reason a file is rejected.
`
WEBVTT
00:00:01.000 --> 00:00:04.500 Welcome to this tutorial on building your first API.
00:00:05.000 --> 00:00:08.200
We will start with the basics
and work our way up.
`
ASS, the SubStation Alpha format, carries fonts, colours and positions and is what fan subtitlers use. TTML is the XML format broadcasters and Netflix exchange. Unless a contract asks for one of those, deliver SRT to platforms and VTT on your own pages, and keep one master file from which the others are exported.
The content of a subtitle file is the same everywhere; the framing differs by format.
What the translation gets wrong, and it is not the grammar
Machine translation of subtitles is good at plain instructions (click the button in the top right), at technical vocabulary with a fixed equivalent, and at the big languages: Spanish, French, German, Portuguese, Chinese, Japanese. It is poor, and confidently poor, at jokes, wordplay, idiom, slang and anything where a word means one of thirty things depending on context. It also flattens tone. A relaxed, friendly speaker comes out formal in most target languages, and in Japanese and Korean the model has to guess a politeness level and picks one.
The subtitle-specific problems are mechanical and therefore checkable. German runs about thirty percent longer than English, so blocks overflow. Some languages take longer to read for the same content, so timings need stretching after translation, not before. Arabic, Chinese and Thai need a different character width for the line limit. None of these are things the model will warn you about.
The approach I would take is the one I take with any supplier deliverable: let the machine produce everything, then put a native speaker on the two or three languages that carry most of the audience, and let the rest go out as machine output with a label that says so. Reviewing every language defeats the point; reviewing none of them is publishing something you cannot read.
Per platform: a file, or burned in
YouTube takes SRT and VTT, one track per language, and your upload replaces its auto-captions, which are usable for English and shaky beyond it. It will also machine-translate your uploaded track into other languages for viewers, which is a reason to upload a good source track even if you never translate anything yourself.
Instagram Reels and TikTok take no subtitle file at all. Either you burn the text into the picture with CapCut, Descript or an editor, or you rely on the platform auto-captions. Burned-in text is part of the design there, so size and placement matter as much as the words.
LinkedIn takes SRT on uploaded video. It autoplays muted in the feed, so the subtitles are the content; keep the lines short.
On your own site, use VTT with the element. Multiple languages are separate tracks, and the text can be styled with CSS.
`html
`
Some generators output blocks in capitals or with random casing. Run the text through the Text Case Converter to sentence case before you split it back into blocks; shouting subtitles read as a mistake, because they are one.
Captions are not subtitles, and the law asks for captions
Subtitles translate speech for someone who can hear the audio and does not know the language. Captions, closed captions or SDH (subtitles for the deaf and hard of hearing), transcribe everything that can be heard for someone who cannot hear it: the dialogue, plus [door slams], [upbeat jazz], and who is speaking. The two are made from the same transcript and they are not the same file.
Accessibility law is about captions. In the United States, broadcast rules and the ADA (Americans with Disabilities Act) reach a growing share of online video. In the EU, the European Accessibility Act has applied since June 2025 to the services it covers, and audiovisual media are among them. In the UK, Ofcom regulates broadcast captions and the Equality Act covers the web. Whether any of that binds a particular channel is a question for a lawyer; the audience argument does not need one. By the WHO's count about one person in five lives with some hearing loss, and a larger group watches on mute on a train.
The recognition models are getting better at non-speech sound, and this is still where a human pass earns its time. The model does not know that the door closing mattered to the plot.
The workflow I would run every week
For anyone publishing regularly, the process is worth more than any one tool in it. This is the one I would set up.
Record for the transcriber: a decent microphone, a quiet room, clear speech. Every error you avoid here is one you do not correct in ten languages. Run the audio through Whisper locally or a cloud service by the minute. Then review the English transcript, five to ten minutes per ten minutes of video, fixing names, terms and the places the model guessed, because every one of those propagates into every translation. Segment into blocks of at most two lines and about forty-two characters per line at natural pauses, if the tool has not done it. Translate with DeepL, Google Translate or a subtitle-specific tool, in the order your analytics say the audience speaks. Put a native speaker on the top two or three languages. Export SRT or VTT per language and upload.
Before upload, three checks, and they are checks on the file, not on the video. Every block under twenty characters per second, in every language. No block longer than seven seconds and none shorter than one. And the last timestamp within the video length, which catches the file that belongs to a different cut. I would rather find the file mismatch with a rule than have a viewer find it with the wrong words on screen.
For anyone publishing regularly, the process is worth more than any one tool in it.
FAQ
How accurate is AI transcription compared with a person?
On clear single-speaker audio the model makes a few errors per hundred words and a professional makes almost none. On crosstalk, accents, noise and jargon the model's rate climbs fast. The gap closes a little every year; the review step does not go away.
Which languages translate best?
The ones with the most training data: Spanish, French, German, Portuguese, Simplified Chinese, Japanese, Korean. Quality drops for languages with less material, Swahili or Tagalog for example, and for regional dialects. If those are your audience, the native-speaker review moves from optional to required.
Burn the subtitles in or upload a file?
A file wherever the platform accepts one. Viewers can switch it off, pick a language and change the size. Burn text in only where there is no alternative, on Reels and TikTok, or when the text is a design element.
What about a video with two languages in it?
Whisper detects a switch and transcribes both, with a dip in accuracy around the switch. For the subtitle file, one track that carries both languages is simpler for the viewer than two tracks with gaps.
AI Chatbot or Search Engine: Which One I Open, and the Question That Decides It
A chatbot when I need something made from the answer, a search engine when I need the source itself. Where each one fails, how I check a chatbot's claims before I act on them, and what changed now that every chatbot can search.
AI Meeting Notes: What the Transcription Gets Right, What the Summary Misses, and the One Line You Say Yourself
Transcription is at 95 percent on clean audio and the summaries are readable. The decision still goes missing. Which tools join the call and which take a file, how to get a transcript worth keeping, why action items need a name and a date said out loud, and what to settle about consent first.
Free AI Grammar Checkers: How They Work and Which to Use
Compare free grammar checking tools, understand how AI grammar checking works vs rule-based approaches, and find the right tool for your writing needs.
Handwriting OCR in 2026: What It Reads, What It Guesses
Printed text is a solved problem for OCR. Handwriting is not, and the failure is quiet: the engine returns a confident wrong word. How the pipeline works, what accuracy to expect per writing style, which engines to use for which job, and how to catch the guesses.