Blog

Meeting Captions Hear the Room. They Never See the Slide.

November 13, 2026  ·  Tablingo

A caption has one input: the microphone. Half of a meeting happens on a shared screen, and the two channels were never connected — which is why numbers and names fail exactly where they were written down.

The presenter shares a slide. On it: a client’s name, spelled the way the client spells it, and a figure with a decimal point in it. Two seconds later the caption line arrives, and both are wrong.

You look at the slide, then back at the caption. The correct spelling has been on screen the whole time. The caption did not read it.

That is not an accuracy problem. It is a wiring problem.

A meeting has two channels. The caption plugs into one.

Everything a caption knows arrives through the audio mix — the microphones, the shared system sound, whatever the platform hands its speech recogniser. The screen is a second channel carrying the same conversation: the figure, the diagram, the code, the name. Nothing connects it to the captions.

Two different things get called “the screen”, and the difference matters. There is the video of the shared screen, which a recording keeps faithfully, pixel for pixel. And there is the text on that screen, which is only data if something reads it. A recording of a slide is a picture of words, not words. You can rewind it. You cannot search it, quote it, or paste it into a document.

Zoom does read the slide. It files the text somewhere else.

Zoom is the one platform that reads screen content while a meeting runs, and its documentation is unusually precise about where that text ends up. Its comparison page credits AI Companion with OCR that extracts readable text from screen-share content during meetings, then adds the parenthesis that matters: supported only for the meeting summary feature.

The data-handling page for those features lists the same three inputs: the speech-to-text transcript, always; text read from shared screens, optional; in-meeting chat, optional. All three are allowed to feed the summary. Only the first feeds the caption, and only the first feeds the transcript.

So the same meeting produces three text objects, and they contain three different amounts of the room.

ArtifactWhat it readsTimeline
Live captionAudio onlyYes
TranscriptAudio onlyYes
AI summaryAudio + screenNo

The third row is the interesting one. The slide’s text did get extracted — and it landed in the only one of the three artifacts with no timeline. You can learn from it that a figure was discussed. You cannot learn when, or by whom, or what was said around it, because a summary is a document about the meeting, not a record of it.

Why it cannot simply be pasted into the caption

The obvious repair — read the slide, push its words into the caption — runs into a property of the caption itself. A caption is a shared stream: one line at a time, identical for every participant, bound to a moment in the audio.

A shared screen is not a stream. In Teams, PowerPoint Live lets each participant move through the deck at their own pace and go back to an earlier slide without interrupting the presenter. Two people in the same call can be looking at different slides in the same second. There is no single “what is on screen” to merge into a single line.

Even in the simpler case — a shared window, identical pixels for everyone — the two channels have no common clock. The caption knows a word was spoken at 00:14:32. Nothing knows that a phrase appeared on slide seven at that moment. Slides do not carry timestamps; speech does.

What gets lost is what was written down

The failure is not spread evenly. It concentrates, and it concentrates on exactly the tokens someone bothered to put on a slide.

Speech recognition resolves an ambiguous sound with context: what came before, what usually follows, what the language model expects next. A surname from another language, a product name, an acronym, a figure read aloud as “forty-two point seven” — none of these have context to lean on, so the model guesses, and it guesses plausibly. That is why the errors read smoothly.

The correct spelling was on the slide the entire time. The two systems never share memory: one listens, the other looks, and nothing in the meeting joins them up.

Which produces the specific frustration of a caption that is fluent and wrong. The sentence parses. The name is not the client’s. The figure is off by a digit. The caption was never in a position to know.

What people do instead

Three moves, in the order most people discover them.

  1. Merge the two channels by eye. Read the caption, glance at the slide, hold both in mind. It works, it costs attention, and it collapses as soon as slides change faster than you can look.
  2. Read the summary afterwards. It has the OCR text, which the caption does not. It does not have the timeline, the speaker, or the sentence around it.
  3. Ask the presenter to say it out loud. The only fix that actually works, and the only one that depends on someone else’s cooperation.

None of these is what everybody actually wants: one artifact that contains the whole meeting.

Two checks

  1. Share a slide with one invented word and one figure with a decimal. Watch whether the caption reproduces either. It will not, and it will not hesitate — the line arrives on time, spelled wrong.
  2. Find the screen-reading switch in your platform’s AI settings, then open what it produced. If the extracted text carries no timestamps, you have found the boundary: the platform can read the screen, but not into the stream.

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

Where our half sits

The honest version of our position: we read the audio of the tab, so the screen is exactly as invisible to us as it is to the platform. We do not do OCR, and we are not going to claim otherwise.

What we cover is the channel with no other reader. If the call is in a language you do not speak, the caption is the only thing standing between you and silence — and it is the only channel that arrives while the meeting is still happening. The figure on the slide you can read yourself. The sentence underneath it, you cannot.

Two boundaries, as always. It works inside a browser tab, so a meeting running in a desktop client is out of reach unless the invite offers a browser link. 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, and what survives a recording is here.