Transcribing your meetings without taking chances with your data.

How Beluga meeting transcription works: an announced recording, audio sent to Mistral AI (European Union) only if the host asks for it, with no retention.

March 18, 2026 · The Beluga team · Product

When you record a Beluga meeting and ask for a transcript, something useful happens: a few minutes later, a document appears in Docs with meeting notes and the full transcript, speaker by speaker. Here is how it works, with no grey areas.

Exactly what happens

  1. The host starts the recording and ticks “Transcribe” if they want a transcript. Every participant is notified: it is deliberately impossible to do discreetly.
  2. When the meeting ends, the recording is saved to Drive, on our servers, in Québec.
  3. Only the audio track is sent to Mistral AI, in the European Union, for transcription with speaker identification. Mistral AI does not keep it after processing and does not use it to train its models.
  4. The transcript comes back and becomes a Docs document, with the meeting notes.

Yes, the audio leaves our servers, and we tell you so

Transcription is not the only data that leaves Québec: payments (Stripe), bot protection at sign-up (hCaptcha) and Convo notifications in Element X (Element) do too, and each one went through an assessment. This one is controlled: it only happens if the host asks for a recording to be transcribed, it is covered by a privacy impact assessment (Law 25), and the audio is processed in the European Union, under the GDPR, without being kept there. Meetings that are not recorded or not transcribed never leave our servers.

Why not host the model ourselves?

Because running a transcription model well requires specialized hardware we do not have yet. We prefer an acknowledged, documented dependency over a false promise of running everything locally. Transcription hosted in Québec is under study, with no date; we will announce it here if it happens.

All posts