Back to Blog
Industry Insight•2026-11-16•12 min read

Real-Time Translation vs Transcription

Real-Time Translation vs Transcription
TL
Team Laxis
Laxis Team @ Laxis

A supplier on the call says something in Portuguese. Your meeting tool catches it flawlessly — every word spelled right, timestamped, attributed to the correct speaker. And you still have no idea what was just agreed.

That moment is the whole distinction, and it's why two features that read almost identically on a pricing page behave nothing alike in a room. One changes the form of a message. The other changes its language. Paying for the first when what you needed was the second is a mistake teams make about once per international expansion.

Four jobs, and two questions that sort them

Every product in this space is doing one of four things, and you can tell which by asking two questions: does the language change, and does the output have to exist while people are still talking?

Transcription keeps the language and changes the form — speech becomes text a reader of that same language could already have followed by ear. That's it. The mechanics of how it's produced, what it costs and where humans still beat machines are a subject of their own, and we've covered how transcription actually works separately.

Translation keeps the form and changes the language. Written source in, written target out. It usually lands on text that already exists, which is why translation and transcription so often run in sequence rather than in competition.

Interpretation changes the language and stays in speech, live, delivered by a person. Different training, different professional body, different bill.

Captioning and subtitling aren't a fifth act at all — they're a delivery format. Text timed to a video track. It can carry transcription (same language) or translation (a different one), which is exactly why the words get used interchangeably by people who mean opposite things. If you're deciding between file types and conventions, how captions and subtitles differ covers the document side.

ModeLanguage changes?Happens live?What you end up withProduced byTypical use
TranscriptionNoEither — live or after the factA searchable text record in the original languageSoftware, a person, or bothMeeting records, interviews, compliance archives
TranslationYesUsually after the factThe same document readable in another languageSoftware, or a human translatorContracts, documentation, published content
Real-time translationYesYes, by definitionTranslated text appearing seconds behind the speakerSoftware, almost alwaysCross-language calls, live events, support chat
InterpretationYesYesSpoken meaning, in the moment; nothing stored by defaultA qualified personNegotiations, courts, clinics, diplomacy
Captioning / subtitlingEitherEitherTimed text over video, sized to fit on screenSoftware, human editors, or a mixAccessibility, sound-off viewing, foreign releases

"Real time" is a claim about latency, not about effort

Vendors apply the phrase to anything that isn't an overnight batch job, which makes it close to meaningless without a number attached. So here's the number the phrase has to beat.

Conversation runs on a much tighter clock than people realise. When linguists measured the gaps between turns in a worldwide sample of ten languages, the most common gap in every one of them fell between zero and 200 milliseconds, and the average across the whole dataset was about 208 — roughly the time it takes to say a single syllable (Stivers and colleagues, 2009). Turn-taking isn't polite pausing; it's a race, and everyone in a conversation is predicting the end of your sentence so they can start theirs.

Set live translation against that and the tiers fall out on their own:

  • Under a second. Fast enough to be part of the conversation. You can read a line and still take the turn it opens.
  • One to three seconds. The realistic band for translated text today, and genuinely useful. You'll be a beat behind, which is roughly where a second-language listener already is.
  • Five seconds and up. No longer live in any meaningful sense. It's a transcript arriving in instalments, and by the time you've read a line the group has voted.

What pushes translation toward the slower end is structural, not a matter of buying servers. A recognition system can publish a word as soon as it's confident. A translation engine can't, because it needs enough of a sentence to know what the sentence is doing — and in German or Japanese the verb that settles the meaning may not turn up until the end. So it either waits or commits early and revises. Both cost something. Why live capture is a harder engineering problem than processing a finished recording is covered in our explainer on how speech-to-text technology works.

If you're the one being translated, finish your sentences. The single biggest thing a speaker can do for a live translation layer is close each thought cleanly before starting the next. Trailing clauses, sentences abandoned halfway and "well, anyway…" all leave the engine holding an incomplete unit it has to guess at.

Interpreters handle this by anticipating where a sentence is going. Software doesn't anticipate — it waits. Short, complete sentences with the verb early will do more for quality than any setting in the app.

Adding a language adds a second place to be wrong

Many live translation products on the market today are a cascade: recognise the speech, then translate the resulting text. Two models, in series, each with its own failure modes.

The arithmetic is unforgiving. Two stages that are each right 95% of the time are right together about 90% of the time — illustrative rather than benchmarked, but the direction is what matters. Accuracy doesn't average across a pipeline. It multiplies.

The more interesting problem is that stage two has no idea stage one made a mistake. A translation model is a fluency machine. Hand it a misheard proper noun or a dropped negative and it will render the error into perfectly grammatical Spanish without a flicker of doubt. Nothing in the output looks damaged, which is precisely what makes it dangerous.

And translation errors don't feel like recognition errors. A misheard word is local — obvious, repairable from context. A translation slip is semantic: a negation flipped, a modal hardened from "we could" into "we will," a pronoun assigned a gender the source never specified. "Let's table it" means postpone in American English and discuss now in British English, and no amount of audio quality helps with that.

Quality is even measured differently. Machine translation gets scored by comparing output against a reference translation — BLEU is the long-standing metric, with learned scores like COMET tracking human judgment more closely — and none of it tells you whether the one line that mattered in your meeting came out right. There's a quieter research track that skips the middle entirely: models trained to go straight from audio in one language to text in another, with no intermediate transcript to corrupt. Fewer stages, fewer places to compound. Plenty of shipping products still cascade.

The ten-second test for whether a tool really translates live. Start a meeting, say two sentences in one language, and watch the screen while you say the third. If a version in your target language has appeared before you finish, that's live translation. If nothing shows up until the call ends and the summary lands, it's post-processing — useful, but it won't help anyone take part.

Run it on a real call, not a demo recording. Demos are chosen for clean audio and one speaker at a time, the condition every tool is good at.

Why a tool can be superb at transcription and useless at live translation

This trips people up constantly, and it isn't a quality gap between vendors. It's a different product underneath.

Excellent transcription can be done in batch: wait for the call to end, process the recording with the benefit of hindsight, take as long as you like. Cheaper to run, easier to make accurate, and it's what most meeting note-takers are optimised for — their whole design assumes you followed the conversation live and just want a record of it afterwards.

Live translation breaks that assumption in three places at once. You hold an open connection for the length of the call. You chop the incoming stream at points where a translation makes sense, which is a linguistic decision rather than a timer. And you pay for a second model on every chunk, for every target language in the room. Holding a stream open costs more to run than processing the same audio once at the end; add a translation pass on top and the economics explain a lot of product roadmaps.

Which is why "does your tool translate?" so often gets a hedged answer. Plenty of tools will translate a finished transcript on request, and that's a genuinely different product from side-by-side text as people speak. We checked the popular note-takers against exactly this question in which AI note-takers actually translate, tool by tool, and the answers are less uniform than the marketing suggests.

Captions vanish. That's what captions are for.

A live caption is a rendering, not a record. It exists to be read once and pushed off screen by the next line, and most conferencing platforms treat translated captions exactly that way: present during the meeting, gone when it ends.

For a status call, fine. For anything with money or a deadline attached, the question that matters is what exists a week later — and there are three possible answers.

Nothing. Translation happened, comprehension happened, and nothing was kept. Whoever wasn't on the call gets a summary written by someone working from memory.

The original-language transcript. Better, and it's what most tools produce. You can search it and quote it, provided you read the language it's in. If you don't, you're back where you started with an extra step.

A translated record. Both versions kept, with the summary and action items carried through in the language each person works in. This is the version that settles arguments. Six weeks on, the dispute is never what the audio sounded like — it's whether the commitment was to deliver by the 14th or review by the 14th, and the document that resolves it has to be readable by both sides.

That third outcome is the practical case for checking what a tool keeps before you buy it, and for running a meeting assistant that translates live alongside whatever captions your platform already puts on screen. Laxis, for one, prints a translation beneath every transcript line while people are still talking (a Premium feature, in 60+ languages), and once the call ends you keep the original transcript, a summary and the next steps in one place. As with any tool, test what is actually stored after a real call before you rely on it. And whoever presses record owns the job of telling participants and getting consent where required; our guide to recording consent across states and countries is a starting point, not legal advice.

See it on a real call. Turn on Translate in the Laxis desktop app and every transcript line gets its translation directly below it, in any of 60+ languages, while the conversation is still going.

14-day Premium trial, no credit card. Try Laxis free

Interpretation isn't a premium tier of the same software

Worth being precise about the third mode, because "we'll just use the AI" is a decision people make without knowing what they're declining.

A simultaneous interpreter listens and speaks at once, running a couple of seconds behind, and managing that lag while keeping the meaning intact is most of the skill. Consecutive interpreting takes the other route: the speaker stops, the interpreter renders the passage, and the meeting moves at half speed in exchange for far higher fidelity.

What no engine does is the part that isn't linguistic. An interpreter notices a phrase is ambiguous and asks. They adjust register when a literal rendering would land as rude. They carry professional liability for getting it right, and in regulated settings that liability is the actual product you're buying.

So the line is about consequence, not difficulty. Weekly standups, project reviews, supplier check-ins, internal training: software handles these and nobody misses the interpreter. A works-council negotiation, a clinical consultation, a deposition, an immigration interview — book a person. Machine translation is a safety net there, not a substitute, and it's excellent at the ninety per cent of conversations where approximately right is fine.

Working out which one you actually need

Four questions, in order. Most teams can stop after two.

  1. Does everyone on the call follow the spoken language comfortably? If yes, you need transcription and nothing else, and paying for a translation layer buys you a feature nobody will switch on.
  2. Does someone need help during the conversation, or only after it? This is the expensive fork. Catching up afterwards is solved by a translated transcript, which is cheap and can be produced at leisure. Taking part during the call needs the live version, with everything that implies about latency and price.
  3. Does the output need to be spoken? Almost never, in business. Synthesised translated speech is improving fast, but text is quieter, skimmable, doesn't collide with the original audio, and leaves a record. Spoken output matters for events and site visits, where nobody can watch a screen.
  4. What happens if one sentence is wrong? If the answer involves a regulator, a court, a patient or a signature, that's the human-interpreter threshold and no accuracy figure changes it.

Most teams over-buy on live and under-buy on the record. Live translation demos well, so it's the feature that gets evaluated. The translated record is what people quietly rely on for months afterwards, and it's what's missing when someone needs to check what was agreed.

Translate the call, and keep a record of it. Read the conversation in your own language while it's happening on Zoom, Google Meet or Microsoft Teams, then come back to the transcript and summary afterwards. Live translation is part of Laxis Premium. 14-day Premium trial, no credit card. Download Laxis

The bottom line

These two capabilities are converging into single products, which is good for buyers and bad for clarity. A pricing page that says "translation" can now mean live side-by-side text, a caption track that evaporates, a summary produced twenty minutes after the call, or a button that runs a finished transcript through an engine on request. Four products, one word.

The distinction worth carrying around isn't technical. Transcription is about memory — making sure the conversation still exists tomorrow. Translation is about access — making sure everyone was actually in it today. A team can be excellent at the first and quietly terrible at the second, and the tell is always the same: someone on every call who nods a lot and never disagrees.

Frequently asked questions

What's the difference between translation and transcription?

Transcription changes the form of a message without changing its language, turning spoken audio into written words that anyone who speaks that language could already understand. Translation changes the language itself, carrying meaning from a source language into a target one. Anybody who could follow the original conversation gains nothing from a translation, and everything from a transcript.

What does real-time translation actually mean?

Real-time translation means the translated version appears while the conversation is still going, close enough behind the speaker that a listener can use it to take part rather than to review later. In practice that means roughly one to three seconds of lag. Anything that only produces a translated version once the call has ended is post-processing, however fast.

Is real-time translation the same as interpretation?

No. Interpreting is performed by a qualified person, delivered as speech, and works at the level of intended meaning, so an interpreter can flag an ambiguity, adjust register, or ask a speaker what they meant. Real-time translation software produces text automatically and has no way to query anything. They overlap in timing and nothing else.

Why does live translation lag behind the speaker?

Because a translation engine needs a big enough chunk of a sentence before it can commit. Word order differs between languages, so a system translating German or Japanese may have to wait for a verb that arrives at the very end. That waiting period sits on top of whatever delay the speech recognition stage already introduced.

How accurate is machine translation in meetings?

Good enough for routine business conversation, unreliable at the edges. Two stages run in sequence, so recognition mistakes are handed to the translator as if they were correct, and the output stays fluent while the meaning drifts. Idiom, in-house shorthand, product names and people talking over one another are where it breaks most visibly.

What is the difference between translated captions and a translated transcript?

Translated captions are a live rendering meant to be read once and replaced by the next line, and most platforms keep nothing once the call ends. A translated transcript is a stored document you can search, quote and revisit weeks later. Captions solve comprehension during the meeting; a transcript solves every argument that comes after it.

Do I need transcription or translation for my international meetings?

If everyone on the call follows the spoken language comfortably, plain transcription is enough and translation adds cost without adding comprehension. Add a translation layer when at least one person is working in a second language, and add it live only if that person needs to contribute during the call rather than catch up afterwards.

Related Product

Laxis Meeting Assistant

Automate your notes today.

  • Bot-free notes for Zoom, Google Meet, Teams and in-person meetings
  • AI summaries, action items and follow-ups in 60+ languages
  • Mac, Windows, iPhone and Android, with a free plan

Start for free, no credit card required.

Learn More ▷