Can AI summarize just the last few minutes of a meeting, not the whole call?

Yes. AI can summarize just the last two or five minutes of a meeting while it’s still going, and the tools that do it fall into two groups. Platform assistants (Zoom AI Companion, Copilot in Teams, Gemini in Meet, and Webex’s Catch Me Up) recap recent discussion when you ask, on calls where the host’s plan and admin allow it. A multi-resolution summarizer like Canary keeps several windows running at once, from the last 30 seconds to the full call, so the summary of the last few minutes is already on screen before you need it.

What doesn’t work is getting there from a whole-call summary. The reason explains why zoom levels exist at all.

Why a whole-call summary can’t tell you what you just missed

A summary is a ranking. To condense forty-five minutes into a paragraph, the model has to decide what mattered across all forty-five, so it keeps the decisions, the disagreements, and the next steps. That’s exactly what you want when the meeting ends.

It’s the wrong ranking for the moment someone says your name. What you need then is sorted by recency: the question that was just asked, the context right before it, and who asked it. In a whole-call summary those are the details that get cut, because two minutes out of forty-five rarely rank as the meeting’s most important thing. Often they haven’t been around long enough to matter to the meeting yet.

A whole-call summary is sorted by what mattered to the meeting. The moment you’re called on is sorted by what happened last. Both orderings are useful. They produce different documents, and you can’t cut one out of the other.

That’s also why “just read the end of the summary” doesn’t work. A summary isn’t a chronological transcript, and its last paragraph is its conclusions, not the last two minutes. The difference between a live transcript and a live summary covers the words-versus-meaning half of this. This page covers the time half.

Zooming in isn’t cropping

The way to get a good summary of the last two minutes is to summarize the last two minutes: take only the words spoken in that window and condense those. Trimming the whole-call summary doesn’t work.

That’s how Canary builds each view. It keeps a separate rolling transcript buffer for each time window, trimmed by timestamp, and writes each window’s summary fresh from its own words. So the 2-minute view isn’t a slice of the full-call view. It’s a separate summary. A detail the full-call view had to discard, like a number, a name, or the exact wording of a question, can survive in it, because in two minutes of speech it may be the most important thing anyone said.

Two design consequences follow, and they’re worth checking for in any tool that claims to do this:

• Windows should slide, not chunk. A “last 5 minutes” view trimmed by timestamp always covers the most recent five minutes. A tool that summarizes in fixed blocks (minutes 0 to 5, 5 to 10, and so on) or by detected chapter hands you a latest block that might be one minute old or nine, depending on when you look. Chapters are good for navigating a recording afterward, but they’re weak for the moment you’re called on.
• Short windows should refresh faster. The gist of a whole meeting doesn’t change in ten seconds, but the last thirty seconds do. Canary refreshes the now view every few seconds and the full-call view every couple of minutes. It skips a refresh when nothing new has been said.

Span and detail are two different zoom levels

“Zoom level” gets used for two things that are easy to mix up, and a useful tool gives you both.

• Span is how far back the summary reaches: the last 30 seconds, the last 2 minutes, or the whole call.
• Detail is how compressed it is: one sentence, a paragraph, or two paragraphs.

A one-sentence summary of the whole call and a two-paragraph summary of the last two minutes might be the same length, and they answer completely different questions. A multi-resolution summary lets you move along both axes quickly.

Here’s what each window is for, and what a whole-call summary does with the same stretch of time.

Window	What it covers	The question it answers	When you reach for it	What a whole-call summary does with that stretch
Now	About the last 30 seconds	”What’s being said right now?”	Looking back at the call after reading a message	Usually leaves it out, because it hasn’t mattered to the meeting yet
Last 2 minutes	The current exchange	”What was I just asked, and by whom?”	Your name is called	Squeezes it into a clause, or drops it
Last 5 minutes	The current thread	”What is this part of the discussion about?”	Coming back from a Slack reply, or when the topic shifts	Folds it into whichever agenda item it belongs to, if any
Full call	The meeting so far, as a gist	”What’s this meeting about, and what’s been decided?”	Joining late, or before raising a bigger point	This is what it’s built for

In Canary, each of those windows comes at three detail levels: Short (one sentence), Medium (a paragraph of three to five sentences), and Detailed (two paragraphs with specifics, decisions, and open questions). That makes twelve views. The widget keeps all twelve current in the background and shows the one you’ve picked, so switching is instant and doesn’t send a new request. When you minimize it, it collapses to a small pill showing the one-sentence now line. The resolution matrix is the name for that grid, and meeting situational awareness is what it’s for.

Which tools can summarize the last few minutes during a call

After the call, any tool can summarize any stretch. Paste a transcript into a chatbot and ask about minutes 40 to 45. The hard part is doing it during the call, and the table below scores tools on that alone. Plans and features change often, so check what your own account includes.

Tool	Summarizes a recent stretch during the call?	How you get it	The catch
Zoom AI Companion	On request, where the host’s plan and admin enable it	Type a question in the in-meeting panel	One answer per question; Zoom only; the host’s account decides
Copilot in Teams	On request, with transcription running and a Copilot license	Type a prompt asking for a recap so far	Same shape; Teams only; the host’s tenant decides
Gemini in Google Meet	A summary-so-far catch-up, where Gemini’s notes are running	Built into the note-taking feature	Paid Workspace plan; the host starts notes; Meet only
Webex AI Assistant	On request, through the Catch Me Up button	Press the button	Webex only; plan and admin settings decide
Otter	A live transcript, and an assistant that answers questions on request	Ask in Otter’s panel	The condensed summary comes after the call
Live captions and transcript tools (Tactiq, platform captions)	No, they show the words	Scroll back and read	Two minutes of dialogue is a lot to read while the call keeps going
After-call notetakers (Granola, Fathom, Fireflies)	No, the summary is written after the call	Not during the call	A great record, but it arrives after the moment has passed
Canary	Yes, continuously: now, 2 min, 5 min, and full call	Glance at the widget; there’s nothing to ask	Fixed windows; only covers calls playing on that computer

Two of those rows are worth comparing directly.

Typing a question is better at one thing. A chat panel can answer questions no fixed window can, like “what did Priya say about the launch date?” or “summarize the last fifteen minutes.” If you need an arbitrary span or one specific fact, asking is the right tool, where the host has enabled it. Treat the time span in the answer as approximate unless you’ve tested how that assistant handles one. Platform-native meeting AI explains why whether you have the tool at all depends on who hosts the call.

A standing window is better at the moment itself. When your name is called you have a couple of seconds. Opening a panel, typing, and waiting for an answer uses them up. The answer is also a snapshot: thirty seconds later it’s stale, and you’d have to ask again. A window that’s already current turns “what did I just miss?” into a glance. The live-summary answers for Zoom, Teams, Google Meet, and Webex go through each platform’s version. The best real-time meeting summarizer roundup ranks the whole category.

Why fixed windows, and why these four

A slider that let you pick any span sounds more flexible. It would also ask you to make a decision at the exact moment you have no time for one. Fixed windows make that choice ahead of time, so catching up means one glance at one of four views.

The spans roughly follow the shape of a conversation:

• About 30 seconds is the sentence in progress and the one before it. That’s enough to tell whether someone is still talking to you.
• Two minutes is about how far back a question reaches: what was asked, and the exchange that led to it.
• Five minutes is about the length of a thread: long enough to hold a proposal and the reaction to it.
• The full call is the arc of the meeting.

These are rules of thumb, not laws, and some questions take ten minutes to arrive. If your meetings run differently, Canary’s settings let you add a personal focus to each window, up to 500 characters each, starting from your next meeting. For example, you could ask the 2-minute view to track “questions asked and who asked them,” and the full call to track “decisions and owners.” You can also hide the windows or detail levels you never use.

How to use the windows when your name is called

1. Look at the 2-minute view at Medium detail. That’s where the question and the exchange around it will be.
2. If anything is unclear, confirm the question out loud instead of answering a paraphrase: “Just to make sure I answer the right thing, you’re asking whether we can ship Friday?” That’s a normal thing to say in a meeting, and it also catches any mistakes in the summary.
3. Zoom out to 5 minutes if your answer depends on the thread, and to the full call if you’re about to make a bigger point.
4. Go back to the call. The windows keep rolling while you talk.

This is the zoom-level version of what did I miss in the meeting? For why you drifted in the first place, see why do I zone out in meetings? The other two catch-up moments, stepping away from a call and joining late, each have their own answer.

What Canary does

Canary 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.

Because it reads system audio instead of joining the meeting, the windows work the same on Zoom, Teams, Meet, Webex, or a Slack huddle. Nothing in the meeting has to be switched on by a host. Canary runs as a small always-on-top widget beside the call on macOS, Windows, and Linux, and it doesn’t take focus away from the call. It’s free for 5 meetings a month, and Pro is $15/mo. For how it compares with the platform assistants in the table, see Canary vs Microsoft Copilot and Canary vs Zoom AI Companion. For how it compares with a live transcript, see Canary vs Tactiq.

What it won’t do

• Pick an arbitrary span. The windows are 30 seconds, 2 minutes, 5 minutes, and the full call. For “the last fifteen minutes” or one specific fact, a typed assistant is better, where the host has enabled one.
• Cover what it didn’t hear. Every window starts when Canary starts listening, so if you joined late, the full-call view covers the call since you arrived. The late-join answer covers what can fill that gap.
• Replace the record of a long meeting. The full-call view is a running gist, and on a long call it leans toward the more recent part of the conversation. For something said early in an hour-long meeting, check the transcript after the call.
• Update instantly. The views are built from finished transcript segments, so even the now view runs a few seconds behind the conversation. Streaming transcription and interim transcription results explain why.
• Always know who asked. A summary can only know who spoke from voices separated in one mixed audio stream and from names people say out loud, so “who asked” can be missing or wrong. How AI notetakers know who’s speaking covers why, and how accurate AI meeting notes are covers the rest.

A narrower window is still other people’s words

It’s tempting to think that summarizing the last thirty seconds is lighter than capturing a whole meeting. It isn’t, from the point of view of the people talking. Canary hears the whole call whichever window you’re looking at, and the window only changes what you see. Your colleagues’ words are captured and condensed by software either way, and the window size is a display setting they can’t see.

So the courtesy doesn’t shrink with the span. Tell people at the start of the call that you’re using a live summarizer, the same as you would for any notetaker. How do I tell participants I’m using an AI notetaker? has wording that works, and one-party vs two-party consent explains why the legal rule depends on where people are. And use the short windows to get back into the conversation, not to quote it. A two-minute summary is a paraphrase, so when you respond, confirm the question in your own words rather than repeating what the widget says someone said.