Teams live captions and translation

Live captions are one of Teams' quietly essential features — an accessibility requirement for some, a survival tool in a noisy café for everyone — and they've grown a family: translated captions, live transcription, language interpretation, and now AI voice interpretation. They're not all the same feature, they're not all in the same licence, and they don't all have the same privacy footprint. This guide sorts them out.

Live captions

Every participant can turn on live captions for themselves during a meeting (More → Language and speech → Turn on live captions). Captions are generated in real time from the meeting audio, shown only to the person who enabled them, and not saved. They're per-user: your captions don't appear on anyone else's screen. The spoken language is set per meeting (the organiser or any participant can change it, and it applies to everyone's captions), and Teams supports a few dozen spoken languages. Speaker attribution appears alongside each caption line.

Included in every Teams licence. Accuracy is good for clear speech in a supported language and degrades with accents, crosstalk, jargon, and bad microphones — the same failure modes as any speech recognition. Teams Rooms devices show captions on the front-of-room display when enabled.

Captions aren't a substitute for CART (human real-time captioning) where an accessibility obligation calls for it. Teams supports CART captions via a third-party captioner using a Teams-provided link; that's a meeting option, not a licence feature.

Translated captions

Translated captions show the captions in a language different from the one being spoken — the speaker talks in German, you read English. This is where licensing bites: translated captions require Teams Premium, and Microsoft has extended the entitlement to users with a Microsoft 365 Copilot licence in some scenarios (check the current licensing table — it has changed more than once). Everyone else sees the "upgrade" prompt.

Translated captions are still per-user and still unsaved. Quality is machine translation of machine transcription — two error layers — and it's fine for following along, not for anything contractual. The number of target languages is larger at Premium than the languages offered in town halls at standard tier, which is one of the practical reasons multilingual organisations end up with Premium.

Live transcription

Transcription is the feature that saves text. When someone starts it, Teams generates a running transcript visible to all participants in a side pane and stores it with the meeting afterwards (in the organiser's OneDrive, or SharePoint for channel meetings). The transcript is what powers intelligent recap and Copilot in meetings — no transcript, no Copilot — and it's what shows up in eDiscovery.

Transcription is included in enterprise and business plans, but it's a policy-controlled feature: the meeting policy sets whether transcription is allowed and who can start it. Because captions are ephemeral and transcripts are records, organisations that are relaxed about captions are often careful about transcription. Both can be on; the distinction matters to legal and to works councils, and it's a distinction users don't naturally make.

Language interpretation (human)

Language interpretation lets the organiser assign human interpreters to language pairs; attendees choose a language channel and hear the interpreter at full volume with the original speaker lowered. It's the feature for multilingual formal meetings — assemblies, councils, board meetings with an interpreter in the room. It's included without Premium, configured in Meeting options ahead of time, and it's about audio, not text.

AI voice interpretation

The newest member: the Interpreter agent, which provides real-time spoken translation — an AI voice speaking the translation, optionally in a simulated version of the speaker's own voice. It's a Teams Premium feature and rolled out progressively; treat it as new and confirm its behaviour, language list, and consent handling in your tenant before deploying it for anything that matters. The voice-simulation option in particular is one to discuss with privacy teams: the speaker must opt in, and some organisations will want it disabled by policy regardless.

Privacy and policy summary

| Feature | Saved? | Licence | Policy control | |---|---|---|---| | Live captions | No | Included | Meeting policy (can be disabled) | | Translated captions | No | Premium (or Copilot in some cases) | Meeting policy | | Transcription | Yes | Included | Meeting policy; who may start | | Language interpretation | No (audio) | Included | Meeting options | | AI voice interpretation | No | Premium | Meeting policy; user consent |

Two settings decide most of the governance: whether transcription is allowed (and by whom), and whether captions are allowed at all. Almost nobody should disable captions — it's an accessibility feature — but some regulated environments disable transcription by default and let organisers enable it per meeting. Sensitivity labels for meetings (Premium) can enforce transcription off for a labelled category of meeting.

Captions and transcription both process audio in Microsoft's cloud speech service within the Microsoft 365 boundary; caption text is not retained, transcript text is a file under the organiser's control. Microsoft's Copilot and Teams privacy documentation covers the data handling.

Practical advice

• Turn captions on in meetings with more than a handful of people; recommend it in adoption material. It's the most under-used accessibility feature in the suite.
• Set the meeting's spoken language correctly — a meeting in Swedish with the language set to English produces gibberish.
• Distinguish captions from transcription in user comms. "Teams is recording what I say" is usually a caption user who didn't realise the difference — or a transcription user who did.
• If you have multilingual staff and no Premium, human interpretation plus untranslated captions in the spoken language is a workable, free combination.
• Decide the Interpreter agent's status with privacy before it decides itself.