Talk to Muse: The Voice Mode Guide
Muse can talk back in real time while it keeps working in the background. When talking beats typing, the exact phrases that work by voice, and the one mistake
By Professor Muse11 min read
I brief most of my tasks by talking. Not because I like the sound of my own voice, but because my best task briefs are the ones I would never type. They ramble. They loop back. They add the detail I remembered halfway through. Typed, that would be an essay. Spoken, it takes ninety seconds, and Muse turns it into work before I have hung up.
That is what voice mode is for. Meta announced it at Connect on September 23: real-time voice conversations with Muse, where the agent keeps working in the background while you keep talking. Think of it as a working session where the other participant has hands, not a phone call with a chatbot. This lesson covers when to talk instead of type, the exact phrases that work by voice, and the one mistake I see every new voice user make.
The short version
- Voice mode is a real-time conversation where Muse keeps working on your tasks in the background while you talk.
- You can describe the voice you want in plain words: its pace, its accent, how warm or brisk it sounds.
- Voice wins for brainstorming, long briefs, and check-ins while your hands are busy. Typing wins for anything precise.
- End every voice brief with “repeat it back before you start.” Misheard words are the failure mode, and the background working makes them expensive.
What voice mode is
Start with the piece that makes it different from every voice assistant you have used before. With ChatGPT Voice or Gemini Live, the conversation is the whole product: you talk, it answers. With Muse voice mode, the conversation is the interface to an agent that is doing other things at the same time. You talk through a trip brief while Muse is already checking flights in the background. You brainstorm a post outline while it pulls the research you mentioned thirty seconds ago.
Meta demonstrated it the way it will matter most. On the Connect stage, Mark Zuckerberg spoke to his Muse, named Agrippa, through Meta’s AI glasses: schedule meetings, check on equipment availability, all while he kept walking. The glasses integration is still “in the coming months,” but the demo showed the shape of it: say your agent’s name and ask it to act on something you are looking at, like a concert poster, while it checks the date against your calendar. Fitness creator Nina Marie Daniele showed the same pattern in a 30-day training program through Oakley Meta glasses: workout pacing, nutrition, progress tracking, all spoken.
Underneath, two separate pieces: Realtime Voice is the voice itself, which you can customize by describing it in plain words. Realtime Avatar is an optional face for that voice: a lip-synced animated video avatar, generated from a reference image, with speech and video produced from one shared stream so the voice, lip movement, and expression stay synchronized. Meta reports roughly 870 milliseconds from the end of your turn to the start of its synchronized response. They are not the same thing; you can talk to Muse for a year and never turn the avatar on.
Availability as of this week: voice mode is on the phone apps. Voice replies in WhatsApp have been there for months. Web-based voice calls are in preparation, per Meta’s own builds, and glasses plus the Muse Charm pocket device are still on the way. If you ask Muse about a surface it does not have yet, it will tell you.
When voice beats typing, and when it does not
I talk to Muse in three situations and type everywhere else. The pattern took a month to settle, and it has not changed since.
| Talk | Type |
|---|---|
| Thinking out loud: brainstorming, talking through a decision, dictating a long brief | Anything precise: numbers, addresses, code, exact wording |
| Hands busy: walking, cooking, getting ready, driving (hands-free) | Drafts you need to read word by word before approving |
| Check-ins on work already running: “how’s the research going?” | Quiet rooms, noisy rooms, or anything sensitive said aloud |
| Long conversations where you want to interrupt and redirect mid-thought | Sessions you need a searchable transcript of later |
Two lines in that table deserve more. First, the interruption thing is the real superpower of a realtime voice interface. In chat, you wait for the full reply, then type a correction. In voice, you cut Muse off mid-sentence: “no, hold on, I meant the other client.” The conversation stays at the speed of thought instead of the speed of typing, and long working sessions feel collaborative instead of transactional.
Second, the noisy-room row is a quality note, not a comfort note. Voice mode transcribes what it hears, and transcription degrades in noise. A brief dictated on a windy sidewalk will have more errors than one spoken at your desk. The errors do not announce themselves.
Prompt 1: Design your voice
Before you do real work by voice, spend two minutes on the voice itself. You describe it in plain words, the way you would describe a podcast host you like: pace, accent, warmth. Paste this and adjust the bracketed parts:
Set your voice to [a calm, slightly brisk tone, moderate pace]. Keep your spoken answers short: two or three sentences at a time, then stop and let me respond. When I ask you to do something, confirm it in one sentence and start working.
The second half of that prompt matters more than the first. The default failure of voice assistants is the monologue: a long spoken answer nobody wanted read aloud. Muse will happily narrate a five-paragraph research summary at you. The “two or three sentences, then stop” line is what makes the conversation usable. Short spoken answers, long written ones: ask Muse to put anything detailed in the chat thread so you can read it later, and keep the voice for discussion.
One honest limitation: describing the voice works, but the range of what you get back is Meta’s to decide. “A Scottish lilt” will move the needle; “sound exactly like my favorite actor” will not, and should not. Keep your expectations at character, not imitation.
Prompt 2: The talk-and-work brief
This is the core voice workflow: start something big by talking, let Muse work in the background while the conversation continues. Here is the shape I use. Say it in your own words; the structure is what matters:
I am going to describe a research task. Start working on it as soon as you understand it, and keep talking to me while you work. The deliverable is [a short written comparison I can read later, not a spoken summary]. I will interrupt you with corrections. If you need an approval for anything, ask me out loud and wait.
Why this works: it sets the two-channel pattern explicitly. Voice stays the fast channel for direction and interruption. The chat thread stays the slow channel for the actual deliverable. Without that line, Muse will try to read the deliverable to you, which is miserable for anything longer than a paragraph.
It also handles the permission question. Muse’s Sentinel still gates sensitive actions while you talk: our Sentinel guide explains the approval layer, and a voice session does not bypass it. When a background task hits something that needs your tap, Muse asks out loud and waits. Say the approval plainly: “yes, go ahead” or “no, skip that.” Do not approve by implication. “Sure, whatever, just handle it” is an accident waiting for a transcript, not an approval.
A named setup: The Commute Kickoff
Here is the voice workflow I run every weekday. I call it the Commute Kickoff. It turns a walk into the most productive thirty minutes of my day.
The Commute Kickoff
- Cadence: Every weekday morning, on my walk, before I open a laptop.
- What I say: “Kickoff. Today I need to [the one big task]. Also [the two small ones]. Research the big one while we talk; I want a written brief waiting in the thread when I get back.”
- How it runs: Muse starts the research in the background and talks through the small ones with me. I interrupt, redirect, and answer its questions while walking. By the time I am home, the research brief is in the chat thread, and the small tasks are either done or waiting on my approval.
- What I never do by voice: approve payments, confirm bookings, or dictate anything with account numbers. Those wait until I am at a screen. Sentinel would stop most of them anyway, but I do not want the decision made at talking speed.
- How it pairs with routines: the Commute Kickoff is manual, but it feeds my scheduled routines. The morning briefing routine gives me the day’s shape; the Kickoff decides what the day is for.
Total typing involved: zero, until I sit down and read the brief. The whole thing works because the brief was designed for voice from the start: spoken direction, written deliverable, approvals held for a screen.
Prompt 3: The redirect phrases
Voice conversations need steering phrases, the way meetings need an agenda. These are the four I use constantly. Say them exactly when Muse is mid-answer:
“Pause. Go back to [the part about X].” “Hold that thought. New priority: [Y].” “Stop. Put that in the chat thread instead and keep it under [200 words].” “Summarize where we are in two sentences, then stop.”
The last one is the most useful phrase in voice mode. Long voice sessions drift: you start on a hiring plan and end up discussing interview questions for a different role. “Summarize where we are” is the checkpoint that pulls the thread back. I use it every ten minutes in a long session, the way I would check the time in a meeting.
What can go wrong
Here is the honest failure mode, and it is specific to voice. A misheard brief is worse than a mistyped one.
What happened to me: I dictated a research task on a walk, asking Muse to compare two vendors. It heard the second vendor’s name wrong, a plausible-sounding but different company. In chat, I would have seen the wrong name in my own message and fixed it. By voice, there was nothing to see. Muse worked in the background for twenty minutes and came back with a careful comparison of the wrong vendor against the right one. The work was excellent but useless.
The fix is the repeat-back rule, and it is non-negotiable for any voice brief that launches real work: end the brief with “repeat back the task, the deliverable, and the deadline before you start.” Thirty seconds of confirmation saves twenty minutes of wrong work. For anything with names, numbers, or dates, I also open the chat thread afterward and skim the transcript. Voice mode keeps the transcript in the conversation; use it.
The deeper lesson: background working multiplies the cost of every misunderstanding, because nobody is watching. A wrong brief in a live chat gets corrected in the next message. A wrong brief in a voice session gets corrected twenty minutes later, after the work is done. The repeat-back is the price of the speed.
One more honest note, on the avatar. The Realtime Avatar is genuinely well made: the lip sync is tight, the expressions track the conversation, and Meta’s own raters preferred it to Runway and HeyGen in head-to-head tests. It also does nothing for the quality of the work. A face does not make a brief better or an approval safer. Use the avatar if you enjoy talking to a face. Do not mistake presence for competence.
Start with one walk
Do not redesign your workflow around voice today. Take one walk, or one commute, and dictate one brief: a research question, a brainstorm, a plan you have been putting off writing down. End it with “repeat it back before you start.” Read what lands in the thread when you get back.
If the brief comes back right, you have found the channel for everything you think faster than you type. If it comes back wrong, you have found your first correction, and you know what to do with those: our memory lesson covers turning a correction into a standing instruction. Mine is “always repeat back a voice brief before starting.” I have not had a misheard vendor since.
