Meeting Transcription

Pia can join a Microsoft Teams meeting as a silent attendee and transcribe it live — splitting the transcript into a separate, colour-coded bubble per speaker. All transcription happens locally on your machine.

What “Beta” means for your data here

Beta refers to reliability, not to a weaker privacy standard. Meeting transcription handles your data exactly like the rest of Pia: the audio is transcribed by the speech model on your own machine, the transcript is written to your device, and nothing is uploaded anywhere as part of transcribing. There is no separate cloud transcription service and no “beta telemetry” collecting your meetings.

What being in beta does mean is that the feature may not always work:

Teams only. Zoom, Google Meet and Webex aren’t supported. For those, use Live Transcription instead.
Joining can fail. It drives a real browser against the Teams web client, so a Teams UI change, a lobby that nobody admits, or an unusual meeting policy can leave Pia unable to get in.
Speaker labels can be wrong. One person may be split across two speakers, or two people merged into one. You can rename speakers during and after the meeting.
Behaviour may change between releases while it settles.

The point at which meeting content does leave your device is when you ask for something to be done with it — choosing Summarize with assistant sends the transcript to your configured AI provider, exactly like pasting it into a chat would. Transcribing alone does not.

What you need

A Microsoft Teams meeting link.
A meeting host must admit the assistant from the lobby before it can transcribe.
On first use, Pia downloads a small automation browser once (this needs internet the first time); later meetings reuse it.
A speech-to-text model must be available — see the Speech guide. Optional speaker separation downloads one extra model when you enable it.
An interactive Windows desktop with a working audio device.

While attending, Pia runs a bundled browser and a background automation helper — both ship with Pia, so there’s nothing extra to install.

Joining a meeting

Open the Assistant and click the meeting button in the message bar (tooltip: Join a meeting and transcribe). This opens the Meeting attendee panel.
Paste the Teams meeting link, or drag the Outlook invite onto the panel — “Or drag the Outlook invite here — Pia reads the link out of it.” It works from classic Outlook’s message list and from a saved .msg, .eml or .ics, and Pia picks the join link out of the organizer and dial-in links beside it. You still tick the confirmation and press Join meeting.
Optionally change the Assistant name in the meeting shown to other participants — it defaults to “‹your name›’s assistant” (or “Pia’s assistant” if you haven’t signed in). Pia appends “ - AI notetaker“ to it, so participants see for example “Alex’s assistant - AI notetaker”, and the panel shows you the finished name. A name that would drop that disclosure is refused.
Tick the consent box: “I confirm I’m allowed to have an assistant join and transcribe this meeting.”
Click Join meeting.

The Join meeting button only becomes active once you’ve ticked consent and pasted a valid Teams link.

Pia then works through a few steps, shown under the title:

Status	What’s happening
Preparing browser…	Setting up the automation browser (downloads it on first use).
Joining meeting…	Connecting to the Teams meeting.
Waiting in the lobby…	A meeting host needs to admit the assistant.
Attending — transcribing…	Pia is in the meeting and transcribing live.

Pia always joins with its microphone and camera off.

What the other participants see

Pia does not sneak in. It joins as a visible, named participant, so everyone in the meeting can see it in the Teams participant list — under the assistant name you set at step 3, which defaults to “‹your name›’s assistant”. A host has to admit it from the lobby, so somebody with control of the meeting actively lets it in, and it sits there with camera and microphone off for the whole meeting.

What Pia does not do is announce that it is transcribing. Teams’ own “recording has started” banner does not appear, because Pia isn’t using Teams’ recording — it listens to the meeting audio and transcribes it locally. So participants can see that an assistant is present, but nothing on screen tells them a transcript is being written.

That gap is why the consent tick exists. “I confirm I’m allowed to have an assistant join and transcribe this meeting” is your declaration to Pia, not the other participants’ consent to you — Pia has no way to ask them. Telling the people in the meeting is your responsibility, and in many places it is also a legal requirement: recording or transcribing a conversation without informing the participants is restricted or unlawful in much of Europe and in several US states.

The practical habit: say at the start of the meeting that your assistant is taking a transcript, and give people the chance to object.

During the meeting

Each person’s speech appears as a chat-style bubble, and every speaker gets their own colour so you can follow who said what. New voices are labelled Speaker 1, Speaker 2, and so on in the order they first speak.

Renaming a speaker

To give a speaker a real name:

Click the speaker’s name (with the pencil icon) above any of their bubbles, or
Right-click a bubble and choose Rename speaker…

Type the name and confirm. Pia keeps that speaker’s colour, relabels their earlier bubbles too, and applies the name to everything they say next.

Speaker separation

Splitting the transcript by speaker (“diarization”) is controlled in Settings → Assistant → Meeting.

Identify individual speakers in meetings — the master on/off switch (on by default). The first time you use it, Pia downloads a small speaker-recognition model; a Preparing Model dialog shows the progress. If the model can’t be downloaded, transcription still works — it just won’t split speakers, and your meeting is never blocked.
Hide speaker labels in the transcript — on by default. Telling speakers apart means measuring each participant’s voice, so naming who said which line is something you switch on rather than something Pia does for you. See below.
Smart speaker detection (automatic) — on by default. Pia re-evaluates all voices as the meeting goes on and automatically corrects earlier speaker assignments, so you’ll see labels tidy themselves up live. No tuning needed.

Speaker labels are off until you ask for them

A transcript arrives without names against the lines. Bubbles are still grouped and coloured by voice, so you can see that the speaker changed — only the labels, which are the part that claims an identity, are left off.

Untick Hide speaker labels in the transcript to have Pia name them. That is also the switch to reach for the other way round: when attribution is visibly wrong, an unlabelled transcript serves a summary better than a confidently mislabelled one.

Turn Smart speaker detection off to reveal manual controls:

Setting	What it does
Speaker matching sensitivity	How similar two voices must sound to count as the same person. Higher splits speakers more readily; lower groups more voices together.
Limit number of speakers	Caps how many distinct speakers Pia creates. Set it to the number of participants so one person isn’t split into several. No limit detects as many as it hears.
Minimum speech length for detection	How much uninterrupted speech is needed before Pia identifies a speaker. Longer is more reliable but very short remarks may stay unlabelled.

Changes apply to the next meeting you join, not the one in progress.

After the meeting

When the meeting ends (or you click Leave meeting), you have three options:

Save… — write the transcript to a Markdown (.md) file anywhere you like.
Save to vault… — file it in your memory vault so Pia can draw on it later in chat. See below.
Summarize with assistant — hand the transcript to the Assistant for a summary.

Nothing is saved automatically — the transcript is only kept if you choose one of the save options.

Saving or summarizing covers the whole session, however long the meeting ran. The suggested file name leads with the date — 2026-09-03_meeting.md — and the dialog opens in the working folder of the chat you are in. Saving a second transcript on the same day asks before it overwrites.

Saving to your vault

Save to vault… opens a short form so the transcript arrives as something Pia can actually find again:

Field	What it’s for
Title	What the meeting was about. Defaults to Meeting ‹date›.
Attendees	Comma-separated names. Pre-filled with the participants Pia saw in the meeting.
Tags	Comma-separated, for grouping related meetings.
Project	Optional.
Notes	Context worth keeping alongside the transcript.

The form shows you the exact path it will be Saved as before you confirm. From there the transcript goes through Pia’s normal document pipeline, so it becomes recallable knowledge — see Documents & Auto-Ingest.

Names in the summary

While Pia attends, it also notes the participant list Teams shows. When you choose Summarize with assistant, those names go along with the transcript, so the summary can say “Anna raised the deadline” instead of “Speaker 2 raised the deadline” — even if you never renamed anyone during the meeting.

The same list pre-fills the Attendees field when you save to the vault.

Choosing the browser

Under Settings → Assistant → Meeting → Meeting browser you can pick which browser Pia drives:

Bundled Chromium (default) — downloaded once; the most reliable choice.
System Chrome / System Edge / Use system default — reuse an installed browser; Pia falls back to bundled Chromium if the chosen one is unavailable.

Show the meeting browser window is off by default: the browser joins hidden. Turn it on if you’d like to watch what Pia is doing.

A meeting stays silent on your machine either way — including with the browser window shown. It used to play out of your speakers, which echoed if you were in the same call from the Teams app on the same device.

Privacy

Audio is transcribed locally. Only model and browser downloads use the network.
Voice fingerprints stay in memory for the duration of a single meeting and are wiped afterwards — nothing biometric is written to disk.
Transcript text and speaker names are kept out of Pia’s log files.
Pia joins muted, and (with the browser window hidden) captures the meeting silently.

See the Privacy guide for how Pia handles your data across the app.