Caption scraping

Caption scraping (sometimes called live-caption capture) is a meeting-capture method in which a browser extension reads the live caption text a video platform is already displaying in the meeting tab — rather than recording any audio — and assembles that text into a transcript it can then summarize. Tactiq is the best-known example; several bot-based notetakers also ship an extension as a lighter-weight second mode.

It is one of the four ways an AI meeting notetaker can get hold of a conversation, and the only one that never touches sound at all.

How it works

The platform, not the extension, does the listening:

1. You turn captions on. Google Meet, Zoom, and Teams each run their own speech recognition to draw live captions on screen. That happens whether or not any notetaker is installed.
2. The extension reads the page. It attaches to the meeting tab and watches the caption element in the DOM, collecting each line as it appears — text that is already rendered, not audio it recorded.
3. It keeps the speaker labels. Platforms usually attach a name to each caption line, so the extension inherits attribution for free instead of doing acoustic speaker diarization.
4. It stitches and summarizes. Lines are de-duplicated into a running transcript — captions rewrite themselves as interim results settle — and that text goes to a language model for a summary and action items.

Nothing is recorded on your machine and no audio leaves it through the extension. The scraped text, however, goes to the vendor’s cloud like any other notetaker’s transcript.

What follows from capturing text instead of audio

Caption scraping’s strengths and its limits come from the same fact.

It’s the lightest-weight method. No participant joins the call, so there’s no meeting bot for a host to admit or block, and nothing to install at the operating-system level — just an extension. On a locked-down laptop where driver installs and desktop apps are blocked, that’s often the only thing that will actually install.

It gets speaker attribution the easy way. The platform has already decided who is talking, so the extension doesn’t have to. This is a genuine advantage over methods that hear one mixed stream of audio.

It inherits the platform’s transcription, warts and all. You get exactly the quality the platform’s captions have, and you can’t swap in a better model. Captions are tuned for glanceable on-screen reading rather than for an accurate record: punctuation is thin, low-confidence words are quietly dropped rather than guessed at, and lines can scroll away before an extension catches them. This is part of why AI meeting notes are approximately right rather than court-record right.

Its coverage is narrow and conditional. Caption scraping works only where all three of these hold: the call is in a browser tab, the platform exposes captions, and captions are switched on. Join from a desktop app and there is nothing to read. Some platforms let a host disable captions for everyone. And anything that isn’t a supported meeting — a webinar, a recorded video, a Slack huddle, a call on a phone next to you — is out of scope entirely, because there are no captions to scrape.

Captions are a transcript, not notes

Because captions are verbatim, a caption-scraping tool’s live output is a transcript: the words, as they’re said. Getting the words faster is not the same as knowing what’s going on — the distinction is covered in live transcript vs live summary. Most extensions resolve this the usual way, by writing the summary after the meeting ends from the transcript they’ve collected, which puts them back in the post-meeting column with everyone else.

Google Meet is where this method is at its strongest, since Meet runs in a browser tab and its captions are free and one click away — which is exactly why the live summary question on Meet has a different answer there than on Zoom or Teams.

Where it sits among the four capture methods

• Meeting bot — an automated participant joins through the link. Device-independent, but visible in the attendee list and blockable.
• Platform-native AI — Zoom AI Companion, Teams recap, Meet’s Gemini notes. Best fidelity, but gated by plan, admin, and host.
• Caption scraping — this page. Nothing joins, nothing installs deeply, but it’s bound to the browser tab and to the platform’s captions.
• System audio capture — a desktop app reads the sound the computer is already playing, so it’s indifferent to which app the call is in and needs no captions, no bot, and no plugin.

The last two are both bot-free; they differ in what they’re bound to. An extension is bound to the tab and the platform’s captions; system audio is bound to the machine. The full pipeline behind all four is in how AI meeting notetakers work.

The transparency point

Caption scraping is quiet by design. No bot appears, and the fact that captions are on tells other participants nothing about whether the text is being saved somewhere — turning on captions reads as an accessibility choice, not a recording one. That makes disclosure more your job here, not less: tell participants you’re taking AI notes. Recording and consent rules vary by region and don’t distinguish between capturing audio and capturing the text of what was said.

Canary takes the other bot-free path. It is a real-time, bot-free meeting summarizer: it captures your computer’s system audio (no bot in the call, no plugin) and shows a live, multi-resolution rolling summary — from what’s being said right now to the whole call — so you can catch up the instant your name is called. That means it doesn’t need captions to be on, doesn’t care whether the call is in a tab or a desktop app, and shows meaning rather than more text to read. For the head-to-head with the best-known caption-scraping tool, see Canary vs Tactiq and the Tactiq alternatives roundup, or the complete guide to bot-free meeting notes for the whole landscape.