Live Answers

Live Answers listens to a call, picks the questions out of it, and answers each one from your library with the snippet it came from.

What happens on a call

Speech is transcribed in your browser, on your own computer. Short pieces of that text go to the model that decides what was asked and which snippet answers it. Each question becomes a card: the question in quotes, who asked it, the answer, what it cited, and a score.

The score is measured, not guessed, and it says what it is made of. The line under an answer carries the number, its band — Certain, Solid, Fuzzy, Weak or No idea — and the reason the server wrote for it, naming how the source was certified, how well it supported the answer, whether sources agreed, and how clearly the question was heard. Hovering or focusing the line shows those four parts as points, and the ceiling when one applied.

Certification is one of those parts, which is the point of the whole thing: a certified snippet ranks above one that needs review; a snippet somebody flagged wrong is never used in a new answer, and a card that already quoted it is marked Flagged wrong. On a card that answered nothing, a flagged match is named. See certification.

What is never recorded

No audio, anywhere. Nothing is written to disk and nothing is uploaded: frames go from the microphone or the shared tab into the recogniser and are gone. After a call the screen says it in those words — Transcribed on this device. No audio was recorded; this list is what was kept.

What is kept is the questions asked and the answers given, for as long as the workspace's History setting says, or until deleted for a call someone kept. See history.

Note: some places require you to tell the people on a call that it is being transcribed. The product does not make that call for you — it states it on first use, and Announce reads the sentence that tells them into the prompter.

Open it

Press Live Answers — on the document toolbar (under More at the narrowest width) or in the workspace header; on a document, ⌘⇧A opens it too. It opens in a tab of its own, one per workspace, so a second press brings the tab that is already listening forward rather than starting again. It keeps listening while you work in another tab. The live feed stays put while it runs; a row under Past calls is a link to that call's own screen.

If an admin has turned the feature off, the page says Live Answers is off for this workspace. and names who can turn it on. There is no Listen to press rather than one that refuses.

Where to go next

• Setup — what to share, which speech model to load, and the workspace's settings.
• During a call — the cards, their bands, and what each action does.
• History — past calls, Keep, sharing, and the retention window.
• Documentation sources — the sites it may look on when the library has nothing.