Transcription & Live Captions
Your Office AI can transcribe calls on the server in real time — delivering live captions with optional translation during the call, and searchable transcripts afterwards that the AI can use as context, all under an access policy you control.
Choosing a pipeline
The transcription pipeline is configured at the deployment level by whoever operates your Your Office AI instance. There are three options, with transcription off by default:
| Pipeline | What it is | Best for |
|---|---|---|
| None default | Transcription is off. No captions or transcripts are produced. | Organizations that don't need transcription |
| LiveKit Cloud | Transcription runs in LiveKit Cloud and arrives over a webhook. | Fastest to enable — no servers to run |
| Self-hosted | A transcription agent runs in your own infrastructure. | Keeping your own one-on-one voice-assistant sessions on your own servers |
New organizations start with transcription set to None. An admin enables a pipeline from Organization settings. When a pipeline is active, participants are notified that transcription is on.
The transcribers
Self-hosted transcription (Modes A and B below) currently covers your own one-on-one voice sessions with the assistant; transcribing a multi-person workspace call requires the LiveKit Cloud pipeline. All three transcribers deliver their text through the same ingest contract, so captions, saved transcripts, and RAG behave identically regardless of which produced them.
- Mode A — Whisper: a solo-AI agent subscribes to the call audio, runs OpenAI Whisper locally, publishes live captions, and posts segments to the server. It can also sample a shared screen to a vision model, and spoken turns show up as your messages in AI chat.
- Mode B — OpenAI: a LiveKit agent session uses OpenAI for speech-to-text (and optionally the LLM and text-to-speech), delivering segments through the same contract.
- LiveKit Cloud: managed transcription delivered to the server over a webhook.
Live captions during a call
- Join a call
Start or join a call in a workspace, with transcription enabled for your organization.
- Turn on captions
Open the call controls and toggle live captions. The overlay appears with speaker labels and timestamps.
- Choose plain or translated
Switch the overlay between transcript and translation to read what's being said in another language on the fly.
- Decide whether to save
Whether the transcript is saved is governed by your Save transcriptions preference (on by default) and the workspace access policy.
What's in a caption
Because captions come from the server, accuracy is consistent across participants regardless of their device. Each caption line includes:
- Speaker label — the display name of whoever is speaking.
- Translation — an optional translated rendering, when the overlay is set to translate.
Saving transcripts
Each user has a Save transcriptions preference, found in Settings, which is on by default. Workspaces also have their own Save transcriptions toggle for calls in that workspace — both the user's preference and the workspace's toggle must be on for a given call's segments to be persisted so they can be searched and used as AI context later.
Who can see a saved transcript
Saved transcripts are governed by a per-workspace, three-tier access policy, so a transcript is only ever visible to the people it should be:
| Policy | Who can read the transcript |
|---|---|
| Whole workspace | Every member of the workspace. |
| All current participants | Everyone who was in the call. |
| Only active listeners | Only those who were actively listening at the time. |
Saved transcripts are RAG-queryable — the assistant can search and cite them in answers, always filtered by the access policy above so people only ever see what they're allowed to. Try "What did we decide in yesterday's standup?" or "Summarize the budget discussion from last week." Embeddings are generated by a background job, with an optional confidence threshold.
Privacy & consent
Transcription is off until an admin turns it on, participants are notified when it's active, access to saved transcripts is constrained by the policy above, and individual users control whether their transcriptions are saved. Data-retention windows for transcripts are set per workspace on the Privacy tab.
Live captions and saved transcripts come entirely from the server-side transcription pipeline — there is no separate in-app "record" button, and Your Office AI does not produce a video recording of the call.