Dictate instead of typing

Overview

The copilot's input box has a microphone button next to Send. Click it, speak, and your words appear in the box as you say them; click again to stop, correct anything, and send as usual. Dictation changes nothing else about the chat — the org strip, the source toggle and the follow-ups work exactly as before.

Three engines can do the recognising. The button shows which one it is using as a small badge, and the tooltip and settings say in one sentence where your audio goes:

With the engine set to Automatic the best available one is picked in that order. You can pin one in the voice settings.

Before you start

• Dictating through MetaLens360 creates a short dictation session — always ephemeral: the server makes it so, whatever the browser asks for — and therefore needs Record meetings and dictate (every role except Viewer by default). On-device recognition needs no permission at all: nothing leaves the browser.
• Voice features need a secure page (https, or localhost in development) and a microphone the browser may use.
• The first time you pick the MetaLens360 or a Browser cloud engine, a short dialog explains where the audio goes and asks you to confirm. Your acknowledgement is recorded with the workspace's consent version; if an administrator changes the terms you are asked again.

Steps

Dictate a question

1. Open the copilot and click the microphone button (tooltip Dictate (on-device) · hold or Alt+↓ for voice settings, with the engine in use in brackets).
2. If a dialog asks Dictate through MetaLens360? or Send audio to your browser's speech service?, read where the audio goes and click I understand, start — or Not now to keep typing. The dialog names the workspace's current Policy version — when your earlier acknowledgement is out of date, or your last attempt to acknowledge failed, it fetches the current terms first — and your acknowledgement is recorded for that version.
3. Allow the microphone when the browser asks. The button pulses while it listens.
4. Speak. Words arrive in italics as you talk and settle when the engine finalises a phrase; each new guess replaces the previous one, so every word appears once. Common Salesforce terms the engines mishear are corrected as they arrive (sokwell 101 becomes SOQL 101, sales force becomes Salesforce, and LWC, DML, CPQ, Apex, SOSL, Revenue Cloud and MetaLens are spelled the usual way). Keep typing if you like — what you type becomes the new starting point.
5. Click the button again to stop, edit the text, and press Send. Sending the question (Enter or Send) also ends a running dictation, so what you say while the answer loads does not pile up in the next message. While an answer is loading the button cannot start a new dictation, but it stays clickable to stop one.

Talk to the copilot hands-free

1. Open the copilot and click Hands-free under the input box. The chip turns red and reads Hands-free · listening, and the microphone starts. The first time, the same acknowledgement and microphone prompts as above appear.
2. Ask your question out loud. You don't press Enter or Stop: when you finish, the question is sent and the answer starts streaming in.
3. The copilot decides you've finished from what you said, not from a fixed timer. A complete question ("How many flows are in production?") goes after about half a second of quiet. A sentence that ends on and, of, for, um and the like waits about two and a half seconds, and if you carry on, the rest joins the same question. A single word ("okay") is never sent.
4. Keep talking for the next question. The microphone stays on, the box starts empty, and the new question never repeats the one just sent. A question asked while an answer is still arriving waits and goes out right after it.
5. Click Hands-free · listening or the microphone to stop. Closing the copilot also stops it, and it never comes back on by itself after a reload. Text you type is never sent for you; press Enter as usual.

If dictation can't start, hands-free switches off and the error says why. Starting on-device recognition gives up after 2 minutes if the language pack is still downloading, and after 10 seconds if recognition does not start; either way the error names the problem.

See and stop a running dictation from anywhere

• While a dictation is listening or paused, a dark pill at the top of every page reads Dictating (or Paused) with a pulsing red dot (amber while paused), the elapsed time and a red Stop button. One click ends the dictation; it asks nothing and keeps you on the page. While a meeting session is live, the meeting's pill is shown instead.
• If you minimise the chat, its header keeps a small red Dictating (or Paused) chip with Stop — the only red one of the two.
• With the MetaLens360 engine, a connection problem turns the engine badge amber for 30 seconds: retrying means MetaLens360 keeps trying on its own; audio not sent means part of what you said was not transcribed. The tooltip names the problem. Dictation keeps running.

Change the language or the engine

1. Hold the microphone button for half a second, right-click it, or press Alt+Down Arrow while the button has focus, to open Voice settings.
2. Choose the Language — English (US, UK, India, Australia, Canada), Deutsch, Français, Español, Português, Italiano, Nederlands, Svenska, Polski, Hindi, Japanese, Korean and Simplified Chinese are listed. On-device packs exist for a fixed set of languages; the others fall back to the MetaLens360 engine.
3. Under Engine, leave Automatic (the popover says which engine it resolves to) or pick one. An engine that is not available here says why — Turned off by your workspace policy, On-device recognition needs Chrome or Edge 139 or newer, No speech-to-text provider is configured for this workspace.
4. Show words as you speak turns the italic interim text off if you prefer to see only final phrases. Spoken replies is shown greyed with coming later.
5. The same settings, and what this device can do, are on the Settings tab of Overview → Meeting Copilot under Dictation in the chat.

Example

Lee, a developer, is on a train with a flaky connection. He opens the copilot in Chrome, holds the microphone button, sets Engine to On-device and says "which flows in acme-uat were changed in the last seven days". The words appear as he speaks, nothing leaves his laptop, and he presses Enter. Back in the office, Priya prefers the MetaLens360 engine for its punctuation: the first time she picks it a dialog tells her that audio streams to MetaLens360 and is transcribed by the workspace's provider without being stored; she confirms once and is not asked again until the policy changes.

FAQs

Is anything recorded?

MetaLens360 stores no audio, whichever engine you use. Where the text goes depends on the engine:

• On-device: audio and text stay on this device; the text exists only in your input box until you send it.
• Browser cloud: your browser sends the audio to its vendor's speech service, which returns the text into your input box. What the vendor keeps is governed by its terms, which is why this engine is off by default and needs your explicit acknowledgement. MetaLens360 receives neither the audio nor the text until you send it.
• MetaLens360: the audio streams to the workspace's transcription provider and MetaLens360 drops it, and the text comes back into your input box. MetaLens360 runs none of the meeting copilot's AI steps on a dictation — no suggestions, no context refresh, no insight extraction — so the dictated text reaches the workspace's AI provider only if you send it as your question. The server holds the text in memory for two minutes after you stop, so the input box can collect the last words, and never stores it; the dictation session's own record holds none of the text — just its title, settings, duration and counts.

Once you press Send, the question goes to the copilot like any question you type.

Why does the badge say "on-device" for me and "MetaLens" for a colleague?

Engines depend on the browser and the operating system. Chrome and Edge 139+ on a desktop offer an on-device model; Safari, Firefox, older browsers and Android do not, so they use MetaLens360 (or, if the policy allows it, the browser's cloud service).

Why did the first on-device dictation take so long to start?

Chrome downloads a language pack the first time a language is used on this device; the tooltip says the first use downloads a language pack. It can take from a few seconds to a few minutes, and happens once per language.

Can I dictate into other fields of MetaLens360?

Today the microphone is on the copilot's input box. Live meeting capture — a transcript of a whole call with suggestions — is a different feature: see Meeting Copilot.

Troubleshooting

The microphone button is greyed out

Hover it: the tooltip gives the reason. Usually the page is not secure, the browser has no speech recognition and the workspace has no speech provider, or your role lacks Record meetings and dictate. The Settings tab of Meeting Copilot shows the same reasons under This device.

"Microphone access was not allowed"

Allow the microphone for this site in the address bar and try again. On macOS also check that the browser itself is allowed to use the microphone in System Settings.

"No microphone was found or it is in use"

Close other applications that may hold the microphone (a meeting client, a recorder), check the device in the operating system and retry.

"On-device recognition is not available for this language here" or "This language is not available for this engine"

Pick a language that has an on-device pack, or switch the engine to MetaLens360.

"The browser could not reach its speech service"

Browser cloud speech needs a network path to the vendor. The MetaLens360 engine and on-device recognition work without it.

"Dictation keeps stopping"

The browser engine is not holding a session (some browsers restart recognition after every phrase, especially on Android). Switch to the MetaLens360 engine.

"Your acknowledgement is out of date; please confirm again"

The workspace policy changed what people are told. Click the microphone again and confirm the new text.

"Could not record your acknowledgement."

The message continues with the server's own reason, for example The consent terms changed; please read and accept the current version. Click the microphone again: after any failed acknowledgement — your first one included — the next click fetches the current terms before the dialog opens, so the new acknowledgement is recorded for the current Policy version.

Related articles

• Use the AI assistant
• Meeting Copilot — private suggestions while you talk
• Meeting privacy, consent and retention
• Roles and permissions