How On-Device Meeting Transcription Works — and Why It Matters

"On-device" gets used loosely, so it’s worth being precise. On-device meeting transcription means the entire pipeline — capturing the audio, converting speech to text, attributing speakers, and summarizing — runs on your own computer. No audio is uploaded, no third-party server sees the conversation, and nothing depends on a network connection. It’s a different model from "encrypted in transit," and the difference is the whole point.

What actually happens on your machine

• Capture. The app records the audio your computer is already playing — the same sound you hear on the call. Nothing "joins" the meeting; there’s no bot and no calendar connection.
• Transcription. A speech-to-text model runs locally and turns the audio into text. Modern local models are good enough that accuracy is no longer a reason to send audio to the cloud.
• Diarization. The transcript is split by speaker, so you get "who said what," not a wall of text.
• Summarization. Decisions, action items, and open questions are extracted — again, locally.

Why the location of the compute matters

When transcription happens in the cloud, your meetings sit on someone else’s infrastructure. Even with good intentions and encryption, that creates a list of questions you can’t fully answer: who can access the audio, how long it’s retained, whether it trains a model, and what happens in a breach. On-device transcription makes those questions moot by construction — the audio never leaves, so there’s nothing to access, retain, or leak.

The trade-offs, honestly

Local processing uses your machine’s resources and depends on your hardware. In exchange you get privacy that doesn’t rely on trusting a vendor’s data policy, transcription that works offline, and a compliance story that’s easy to explain: the recording stayed on the device the whole time.

Celeritas is built entirely around this model. Meetings on Zoom, Meet, or Teams are transcribed, diarized, and summarized on your own Mac and turned into ready-to-push tickets — with no cloud upload at any step. For teams handling anything sensitive, that’s not a feature; it’s the reason the tool exists.