Voice conversations

Voice is a mode of the same assistant, not a separate product. Same permissions, same scope, same conversation history — you just speak instead of typing.

What this does

Press the microphone in the assistant composer and the assistant opens a full-screen voice stage: an orb that shows what’s happening, a countdown, and a panel beside it for anything you need to read rather than hear.

The orb is the session made visible:

Orb state	Meaning
Connecting…	Setting the session up. Your microphone isn’t live yet.
Listening	It can hear you.
Thinking	It’s working out an answer.
Speaking	It’s talking.
Microphone off	You muted it.
Connection problem	The connection degraded.

The stage only appears once the microphone is actually live — never on the press — so the full-screen surface never tells you that you’re being recorded slightly before you are.

Before you start

Accept voice processing. It’s an optional consent in your account settings, separate from AI assistance. Speaking sends your recorded voice to a third-party AI service; typing does not, so the two are asked separately and you can accept one without the other.
Allow your browser to use the microphone when prompted.
Your organization needs a model provider whose account includes realtime voice.

Start a conversation

Open the assistant (Ask in the header, or ⌘J).
Choose the microphone — Start a voice conversation.
Wait for Listening, then talk.

While you’re in it:

Mute / Unmute stops and resumes sending audio.
Transcript opens the running written transcript in the side panel.
The timer reads “3:12 elapsed · 6:48 left”. Sessions have a hard server-side length limit; the countdown is so an ending reads as a limit rather than a crash.
End finishes the session.

Every voice session opens with a spoken and on-screen disclosure: “You are speaking with an AI assistant. It can be imperfect — check anything that matters.”

Ask it about your data

A staff voice session can answer questions about your organization out loud — how many learners are enrolled, what’s waiting to be marked, how a course is going. An idle stage suggests things you can ask.

It is read-only, and says so on the stage: “It can look things up for you, but it can’t change anything.” Voice can look something up; it cannot make a change, and this is not a setting.

It answers back quickly — it acknowledges you before it starts looking, and a simple count comes back without a round trip to the model at all. Where a lookup does run long you hear “still checking” rather than silence, with an elapsed hint on screen (“12s — still looking”) so a pause reads as a pause rather than a hang. The answer arrives spoken and in the companion panel, with the evidence behind it — so a figure you were told is a figure you can check.

Every gate that applies to typing applies here, re-checked on the server: your permissions, your organization’s AI budget, rate limits, and consent. A refusal is spoken as a sentence rather than dropping the call.

Learner voice sessions get none of this — no data tool, no lookups.

The panel beside the orb

Some things must be read. The companion panel docks next to the orb without ending the session, and carries whichever of these applies:

Transcript — what has been said so far, labelled You and Assistant.
The answer to a data question, as text, so you can read a number rather than trying to remember it.
What I looked at — the evidence behind a figure it just gave you.
From your documents — a citation from a document attached to the conversation.
Confirm this change — a proposal.

Pressing Esc closes an open panel before it ends the call, so you can’t lose a live session by dismissing a panel.

Changes are never confirmed by voice

A proposal takes the panel and waits for a press. The panel says so: “Changes are confirmed here on screen, never by voice.”

This isn’t caution for its own sake. “Yes” is the single most misheard word in speech recognition, and a misheard yes on a bulk withdrawal is unrecoverable. So the assistant will describe a change out loud, but the Confirm and Not now buttons are the only way it runs — and everything from Ask the assistant still applies: mass changes ask twice, publishing still needs a second approver.

What is recorded, and what is kept

Audio is not stored. It is streamed to the model provider so it can hear you and reply, and not retained by Lurno.
A written transcript is kept, alongside your other assistant conversations — private to you, retained for 180 days, covered by the same data-export rights as everything else in your account.
The consent text you accept says exactly this, and you can withdraw it at any time without losing the typed assistant.

Reference — why voice might not start

Each refusal is a different sentence because each has a different fix.

Message	What to do
Voice isn’t switched on for learners in this organisation.	Learner voice is off by default. An administrator can ask Lurno to enable it.
Before you can use voice, you need to agree to voice processing in your account settings.	Accept the voice-processing consent in your account.
No voice provider is configured for this organisation yet.	No connected provider offers realtime voice.
This organisation’s AI provider account doesn’t include voice.	Your provider key works, but the account isn’t entitled to realtime. Add it with your provider.
This organisation’s voice allowance is used up for now.	The voice budget is exhausted. Budgets are set per organization by Lurno.
Voice isn’t configured on our side yet. We’re on it.	A Lurno-side problem. Nothing for you to change.
Voice is busy right now. Please try again in a moment.	Rate-limited. Retry shortly.
We couldn’t use your microphone.	Your browser blocked microphone access. Allow it in site permissions and retry.
That took too long to start.	The session credential expired before connecting. Start again.
The connection dropped.	Network. Start again.

Troubleshooting

There’s no microphone button. Either your organization has no realtime-capable provider, or you’re a learner in an organization where learner voice hasn’t been enabled.

It asks for consent every time. The consent is recorded per person, per policy version. If the policy text has been updated you’ll be asked again.

The session ends on its own. Sessions have a hard length limit, enforced by the server whether or not the countdown was on screen. Start another.

It heard me say yes and did nothing. Correct. Changes are confirmed with a press, never by voice.

I asked it to change something and it wouldn’t. Voice is read-only. It can look things up and read you the answer; making the change is done by typing, or on the page itself.

It says it’s busy, or that I’ve asked too often. The same rate limits and budget that apply to typing apply to voice, and they’re re-checked on the server. Wait and ask again.

I want the assistant but not the microphone. Decline or withdraw voice processing. AI assistance and voice processing are separate consents, and the typed assistant keeps working.