Blog

Live Captions Leave. The Recording Comes With Its Own.

November 9, 2026  ·  Tablingo

Every meeting platform keeps captions after the call. It is a second pass: a different engine, a shorter language list, and never the pair of lines you were reading while people spoke.

You spend ninety minutes in a call reading two lines at a time — the language of the room on top, yours underneath. It works. You follow the argument, you catch the objection, you know what was decided.

Then you send the recording to the colleague who was asleep at that hour. They open it, scrub to the middle, and ask the question you did not expect: where are the subtitles?

They are in there. They are just not the ones you watched.

The recording is not a container

The instinct is to picture the recording as a box that holds the meeting, captions included. The reality is closer to the opposite. The captions were a rendering of the audio — produced once, for eyes that were present. What gets stored is a second attempt at the same audio, made afterwards, by a different process, for a different purpose.

Each platform draws that line in its own place.

PlatformWhat survivesCaption languages
ZoomA one-language transcript19
Google MeetA Doc, plus a track in the video8
Microsoft TeamsOriginal audio only0

The third column is the one worth remembering. In every case it is shorter than the language list the platform offered you during the call.

Zoom: a transcript, an audio channel, and a dropdown that is not translation

Start with the good news, because Zoom does more than most people assume. A cloud recording carries an audio transcript: generated automatically, embedded in the audio and video files, hidden by default. There is a CC button in the playback view. Press it and the words appear over the picture, synced to the sound.

Two conditions. Transcripts are a Business, Education and Enterprise feature. And the transcript arrives in one language, chosen from nineteen — English, Spanish, French, German, Italian, Simplified Chinese, Russian, Ukrainian, Japanese, Korean, Vietnamese, Dutch, Portuguese, Arabic, Polish, Romanian, Swedish, Turkish, Danish.

Now the part that catches people. Hover over the transcript file in the recordings list and you are offered Change Language. It reads like translation. Zoom’s own documentation describes what actually happens: the recording is reanalysed and transcribed in the newly selected language.

Reanalysed. Not translated. The engine returns to the audio and listens again, this time expecting something else. If the room spoke Japanese and you select English, you do not receive English. You receive a worse Japanese — the same sound, heard by a listener who was told to expect a different language. The dropdown changes the listener, not the language.

And the interpretation channels? Here Zoom has improved, and the improvement tells you what kind of object a recording is. Since late 2022, a cloud recording of an interpreted meeting captures the interpretation audio as well, as one separate audio file per language. Every channel you could have selected during the call does survive it.

As audio. There is no text in those files, no caption track to go with them, and no per-language transcript to match. Somebody still has to listen, or feed each file to a transcription service, before a French colleague can read a Japanese meeting that was already being interpreted into French, live, for two hours.

The interpretation was a performance. What got archived was the tape.

Meet: three texts from one meeting, none of them yours

Google Meet can produce three separate text artifacts from a single call, and the surprise is that none of them is the pair of lines you were reading.

So one meeting yields a Doc in eight languages, an embedded track in whatever language was chosen at record time, and a Drive track in twenty-nine. Three texts, three engines. The bilingual pair you were reading is not among them, because it was drawn on your screen and never written down.

Teams: the translated layer is not saved at all

Teams is the shortest case. Translated captions and translated transcripts exist while the meeting runs and are not accessible once it ends. The recording keeps the original audio. There is no second pass to compare, because the first one was never written to disk. Who is licensed to switch it on is a separate article; here it is one line in the table and nothing more.

Why nobody just saves what was on screen

At this point the temptation is to call it an oversight. It is not. The two objects have opposite jobs, and the design follows from that.

A live caption has one obligation: arrive before the sentence stops being useful. Everything else is sacrificed to it. It cannot wait for the end of a thought, because by then the room has moved on. It cannot be corrected, because the viewer already read it. It cannot label speakers, because it does not yet know who is talking. Text is emitted, displayed and dropped. The bet is that being early beats being right.

A recorded caption has the opposite obligation: survive being quoted. It can wait. It can be listened to again, corrected, split, labelled, indexed, searched, exported and re-embedded. Its quality is judged months later, by someone copying a sentence into a document.

Same audio, different objective, different text. That is not a fault in the pipeline; it is the pipeline working as specified. It also explains the thing people find most puzzling — why there is no “save caption” button anywhere. The live caption was a rendering. A rendering is not a document, and storing it would produce a file that is both late and wrong: it would carry every error of the fast pass with none of the corrections of the slow one.

The same split outside meetings

This is not a quirk of meeting software. It is how captions behave everywhere.

YouTube draws the line in the same place. Automatic captions on a live stream are not kept; the replay is captioned again from the finished file, and the two versions can differ. On a finished video, auto-translate into more than a hundred languages exists — but only on top of a caption track that already exists. No track, nothing to translate. Auto-dubbing, now on by default for eligible channels in twenty-seven languages, replaces the voice you hear; it does not add a word of text.

Twitch is blunter, because it never had a document layer to lose. Platform automatic captions arrived in 2022 and work during the broadcast; when the stream ends, nothing is written to the VOD. The VOD remains, uncaptioned. Captions survive only if the broadcaster pushed them into the stream itself — a native closed-caption track from the encoder, which persists into the VOD and into the mobile apps; or text burned into the picture, which persists because it is no longer text. An extension overlay, the easiest of the three to install, survives nothing.

Audio-only is the end of that line. A podcast has no picture, so a caption has nowhere to sit, which is why the platforms there built transcripts instead — a different object, with a different shape and a different set of limits.

What to do, if you are the one who needs the words

Three moves, ordered by what they cost you.

  1. Decide during the call. Nothing here is retroactive. A transcript exists only if someone with the rights pressed start while the meeting was running, and it lands in the organiser’s Drive, shared with the host, the co-hosts and whoever started it. If you were a guest, you may be looking at a file you are not allowed to open.
  2. Expect the transcript in the language of the room. Every platform transcribes what it hears, and none of them translates the file afterwards — Zoom’s dropdown re-listens, it does not translate. If the call was in a language you do not read, a transcript is not a solution. It is the same problem in a different font.
  3. Budget the captioned file as a separate job. If what you need is a video with subtitles in your language, it will be made by feeding the MP4 to a transcription tool and then to a translator. That is a second pass, with its own errors and its own hours. “Can we caption the recording?” — yes, but not by exporting it.

Two checks

  1. Open a recording you already have and press CC in the player. If nothing appears, no captions were written into that file, and no setting will add them now. Drive is the exception worth knowing: it can generate a track afterwards, in its own languages, and it will not match what you saw live.
  2. Find the language dropdown on the transcript. If changing it triggers a reanalysis rather than a translation, you have found the boundary: the platform can listen again, it cannot read across.

Both take a minute and require no trust in any vendor, including us.

Where our half sits

Which brings the honest limit of this piece. We do not produce the file. The extension reads the audio of the tab you already have open and puts the original line above the translation while the call is happening. If what you need is a searchable transcript next week, a transcription service is the right tool, and we are not one.

What we cover is the half with no second chance. The interpreted layer in a call lasts exactly as long as the call does — and as the three platforms above show, what survives it is either audio without text or text without the translation. If nobody was reading it while it happened, no export will give it back.

Two boundaries, plainly. It works inside a browser tab, so a meeting running in a desktop client is out of reach — if the invite has a “join in your browser” link, that is the way in. And it is machine translation: enough to follow a call, not a substitute for a professional interpreter when the words are contractual.

If you want to try ours, it’s called Tablingo — a browser extension that reads the tab you are already in and puts the original line above the translation. The platform-by-platform meeting case is here, who controls translation inside a call is here, why every transcript tool waits for the end is here, and the half of a meeting that never reaches any of these texts is here.