Blog

How AI detects questions during a live meeting

The Trippi Cue team · 2026-09-07 · 7 min read · facts checked against the product on 2026-09-07

Detecting a question in live speech is not the punctuation problem it looks like. Recognised speech often arrives without a question mark at all, half of the requests in a real meeting are phrased as instructions, and the hard part is not spotting a question but deciding whose it is. In practice it takes two stages: a cheap local filter that throws away the obviously empty, and a model that judges ownership with the conversation as context.

Why the question mark is useless

Speech recognition adds punctuation as a guess, and some sources give none at all. “So you would keep the old system readable for a week” is a question in the room and a statement on screen. Meanwhile “Walk me through the pricing” is a request that will never carry a question mark in any transcript, ever.

Anything built on punctuation alone therefore misses two large classes of request: unpunctuated questions and imperatives. Both are extremely common in professional calls.

Stage one: a filter that is deliberately generous

The first stage runs on your machine, on every recognised line, and its only job is to decide whether a line is worth asking a model about. It looks for a small number of signals: an opening interrogative word, an imperative opening, a name that belongs to the user, a question mark when there is one.

It also drops what could never be a request — lines made entirely of “yeah”, “sure”, “makes sense”, and anything too short to carry one.

The design rule here matters more than the word lists: this filter should err towards letting things through. A missed question is a product that stayed silent when you needed it; an extra check is a fraction of a cent. Tightening this stage to save money is the most tempting and most damaging optimisation available.

Stage two: whose question is it?

This is the judgement no keyword list can make, and it needs three things a single line does not contain: the previous lines, the names of the people in the call, and your own name.

With those, a model can separate “Karl, can you share your screen?” from “And how would you handle that?” asked straight after you finished speaking. It can also recognise the rhetorical question — a question in shape, an assertion in function — which is otherwise indistinguishable.

Two smaller decisions belong here too: whether the request is a question (answer it) or a task (say how you would approach it), and how confident the model is that it read the room correctly.

Stage three, the one nobody advertises: the gate on the way back

A model asked to be helpful will be helpful too often. So the answer is checked before it is shown at all:

• Confidence threshold. Below it, nothing appears. A wrong card mid-conversation costs more attention than a missed one.
• Deduplication. Live captions revise themselves, so the same question arrives twice with two words changed. Without merging, you would get two cards for one question.
• Your own speech. A question you asked somebody else must never come back to you as a card.

Enforcing these on arrival rather than in the prompt is the difference between a rule and a request: an instruction the model did not follow is indistinguishable from an instruction that was never given.

What detection still gets wrong

Names that recognition mangles are the most common failure: if your name never lands correctly in the text, the strongest signal is gone and detection leans entirely on context. Overlapping speakers are the second — when two people talk at once, whose turn ended is genuinely ambiguous.

Non-English calls are a third, softer case. Word-list heuristics are language-specific; in a language they do not cover, more of the work falls to the model, which means more calls rather than fewer cards.

This is why any serious implementation keeps a manual path: a field where you can ask the thing directly, skipping every filter. When automation is unsure, a human who is sure should be able to override it in one gesture.