Live Transcription

Live Transcription writes down a conversation as it happens: your microphone and whatever this device is playing. It’s for the calls and conversations that aren’t a Teams meeting — a phone call on speaker, a room, a video call in an app Pia can’t join.

Transcription runs on your machine. The transcript reaches your AI provider only if you choose to: when you save it to the vault, which your provider then works into your knowledge base, or when you ask for a summary.

Starting a Session

Open the Assistant and click the live transcription button in the message bar (tooltip: Live transcription with consent)
Read the Before you start screen
Tick I have informed everyone present and I’m ready to start
Click Start

The Before you start screen names the three purposes Pia uses the transcript for:

transcription on this device;
if you choose, saving it as a file or as a vault note — your AI provider then works a vault note into your knowledge base;
if you choose, a summary by your AI provider, which receives the transcript for it. The summary is saved as a chat together with the transcript and, with sync turned on, synchronised with your account.

After Start, Pia shows Preparing…, Starting…, and finally Listening….

Use Stop to pause and Resume to pick up again in the same session.

If a start fails and you try again, the transcript carries on, and new voices are numbered on from the highest label it already holds. Every voice keeps a label of its own, so renaming or revoking always reaches exactly one person.

This is the part worth understanding, because it changes what actually ends up in the transcript.

Your microphone side is transcribed from the start. Everyone speaking into your microphone appears as me — Pia doesn’t try to tell them apart.

The other side — anything coming out of your speakers — is not transcribed until that person says a consent sentence naming Pia. Until then their speech is only checked for that sentence and then thrown away. It’s never written down.

On that side, Pia only lets text through that it clearly matches to a person who agreed. Speech it can’t match with confidence is dropped, and it doesn’t count as consent either.

The sentence to say is:

My name is [Name] and I accept this recording by Pia.

German and French work too:

Mein Name ist [Name] und ich akzeptiere diese Aufzeichnung durch Pia.

Je m’appelle [Nom] et j’accepte cet enregistrement par Pia.

These are shorter than the wording earlier versions asked for, and the German one has lost its comma — the old phrasing invited a pause in the middle, which split it into two utterances and got it refused. The older wording still works if somebody says it.

Getting the sentence to the other side

Each sentence has a Copy this sentence button, and the one in your language stays at the top of the transcription window for the whole session. Paste it into the meeting chat: that is the only way somebody joining over system audio gets to read it.

A chime confirms each accepted sentence. You hear it; the far end does not.

A refused sentence records which of the four required parts was missing — the part only, never anything anyone said.

Each voice on the other side gets a chip. It starts out with the label Pia gave the voice, such as Speaker 2. Once the sentence is recognised, the chip fills with the accent colour, gets a check mark and — with Name speakers on — shows the name from the sentence. Hover over a chip to see where it stands:

Status	What it means
awaiting consent	Heard, but nothing is being transcribed.
consented	The sentence was recognised — this voice is now being written down.
revoked	Consent was withdrawn. Transcription stops again.

Right-click a speaker to Rename speaker or Revoke consent. Rename speaker needs Name speakers to be on. A new name that another voice in the transcript already carries is refused, including one from an earlier session of the same transcript, so one label always names one person.

Revoke consent stops transcribing that voice again. If the session’s words have already gone somewhere, Pia then opens Copies of this session and lists them:

Exported files, with path and time. Pia cannot reach these files — delete or clean them up yourself.
Vault notes. Delete removes one note, Delete all every note listed, and what a note contributed to your topic pages goes with it. Notes you keep record the revocation in their front matter.
Chats — the session’s summary chats, which you can delete the same way.

If Pia could not check every vault note, the dialog says so, because notes may then be missing from the list.

When Name speakers is off

Name speakers lives in Settings → Assistant → Meeting and is on by default. With it off, the name in a consent sentence doesn’t label the speaker. The consent still counts, and the name is kept only in Pia’s own consent record, not in a saved transcript. The sentence itself stays out of the transcript, and every voice keeps a label such as Speaker 1.

Speaking Time

Click the statistics button in the header for Speaking time — how much of the conversation each speaker has taken. Handy for interviews and one-to-ones.

Saving

When you’re done, you have three options:

Save — write the transcript to a Markdown (.md) file wherever you like.
Save to vault… — file it in your Memory vault with a title, attendees, tags, project and notes, so Pia can recall it later in chat. See Documents & Auto-Ingest.
Save & Summarize — save it and hand it straight to the Assistant for a summary.

Nothing is saved automatically. Close the panel without saving and the transcript is gone.

A vault note stays on this device: the vault is not synchronised with your account.

A file you save and a vault note carry the consent record in their front matter, the block at the top of the Markdown file:

for each speaker who consented, the label Pia recognised the voice under and the label the text shows — which can be a name after renaming — with the time of consent and any revocation;
the version, purposes and language of the notice you were shown, and the id of each session the transcript comes from.

It holds no audio and no voice features. When your AI provider works a vault note into your knowledge base, the front matter goes to the provider along with the note. A summary goes without it.

Privacy

Transcription is local. No audio leaves your device. The transcript does only when you choose: Save & Summarize sends it to your AI provider like any other message, and a note saved with Save to vault… reaches your provider when it is worked into your knowledge base.
Non-consented speech is never written down. It’s checked for the consent sentence and discarded.
Voice fingerprints live in memory only for the session and are wiped afterwards.
A record of each consent is kept so you can show one was given — the sentence, the name, the time — stored under %LOCALAPPDATA%\Pia\ConsentEvidence and protected by Windows. A later revocation is recorded beside the grant, never over it.
That record identifies each speaker by the label Pia detected them under, and so do its file name and a revocation. A name you typed for someone before they agreed never appears there.
Pia keeps that record only while Pia still holds a copy of the transcript: a vault note from the session, a topic page built from such a note, or the session’s summary chat. Once the last of them is gone, the record is deleted at the next clean-up. Without any such copy, it is deleted when the session ends. Pia clears what is due every time it starts, and once a day while it runs.
Files you save are yours to look after. Pia doesn’t track an exported file, so it keeps no record for one — the file carries its own. A note you move out of the vault, or a vault folder you switch away from, counts like an exported file.
If a vault note can’t be read, Pia deletes no record in that run, because the note might still need one. A note that stays unreadable keeps every record until it can be read again.
The 14-day limit now applies only to records left over from Pia versions that predate this rule. Those are still deleted 14 days after their last change; every newer record follows the transcript copies as described above.
A consent log under %LOCALAPPDATA%\Pia\ConsentAudit notes what happened in each session — when it started and ended, when a speaker was detected, gave or withdrew consent, and when unconsented speech was dropped — each with the time and the label Pia detected the speaker under. It holds no transcript text, no consent sentence and no name, and each file is deleted 14 days after its last entry.
Transcript text and speaker names stay out of Pia’s log files.

How this differs from a meeting

The consent gate only works if Pia can tell voices apart, so a few of your meeting settings are deliberately overridden here:

Speaker separation is mandatory. A meeting degrades gracefully if the speaker model can’t be downloaded; a live-transcription session won’t start at all, because “one anonymous speaker” would mean nobody’s consent could be tracked.
Smart speaker detection is off, whatever your meeting setting says. It corrects earlier speaker assignments as it goes, and under a consent gate that could retroactively hand a consented label to someone who never consented.
Your speaker limit is ignored. At a cap, Pia would be forced to fold a new voice into an existing speaker — and that voice could inherit a consent it never gave. Over-splitting one person into two is the safer failure, so that’s what happens instead.

Your speaker matching sensitivity setting still applies.

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

Next Steps

Meeting Transcription — have Pia join a Teams meeting instead
Speech (STT & TTS) — the speech-to-text engine this uses