Muse Voice is live: how to start a voice chat, pick from hundreds of voices, and talk face to face with the avatar

Voice Watch

The short version

Until November 5, talking to Muse meant the staged sightings this site has tracked since late October: a rollout spotted in testing, reported by one outlet chain, with no announcement from Meta. That ended on November 5. Meta published its own post, titled "Give Your Muse a Voice," and the launch is general: voice and expressive avatars, rolling out now in the US, on mobile, web, and Mac.

Two pieces of technology sit underneath, and Meta names both in the coverage of its research work: Muse Realtime Voice and Muse Realtime Avatar. TechCrunch names both technologies in its November 5 report. The voice answers right away and can be interrupted mid-sentence, the way you would interrupt a person. The avatar is a face for that voice. It sits at the top of your conversation, moves its expressions and gestures with the tone of what is being said, and reacts while you are talking, so you can see it following along.

What this launch is not: it is not worldwide, and it is not on glasses or the Charm yet. Meta's announcement also does not say which of Muse's text abilities carry over to voice, which languages it speaks, or what happens to recordings of your voice. Those gaps stay in this post, in their own section, because on launch week the unanswered questions are part of the story.

How to start talking, step by step

Meta gives exactly one starting instruction, and it is short. The steps below use only that instruction and the display and voice choices named by Meta and The Verge. Steps no source has named are listed as gaps after the steps, not invented.

1. Open the Muse app on mobile, on the web, or on Mac.
2. Tap the phone icon in the top-right corner to start talking (Meta).
3. Choose how the conversation looks: voice only, the avatar alongside your conversation, or full screen when you want to talk face to face (Meta).
4. Speak. Meta says you can speak to your Muse as you do with real people, hear responses right away, and interrupt if you need to.
5. To change how Muse sounds, either ask the agent itself for the tone you want, or browse the options in the avatar settings (The Verge, November 5).

There is no separate video button in anything Meta or the coverage describes. The avatar is part of the voice conversation, not a second mode you start on its own. The Verge, which tested the feature on the iOS app on November 5, describes the avatar sitting at the top of the conversation and animating in time with the words it is saying.

What no source describes yet: a wake word, a settings path for turning voice on, or how a session ends. If those exist, they were not in the announcement or the first day's coverage, and this post will add them when a source states them rather than guess at menu names.

Interruption matters more than it sounds. A voice assistant you cannot interrupt is a recording with a microphone. One you can cut off and redirect is a conversation. That is the difference between using voice for a quick question and using it to work through something longer, like the interview rehearsal and presentation practice Meta uses as examples.

The avatar: a face that reacts, and that you can replace

The avatar is the part of this launch that will get the screenshots, so it is worth being precise about what it does and does not do.

It reacts. Meta says that as your Muse speaks, the avatar uses subtle expressions and gestures that match the tone of the conversation, and that it reacts while you talk, so you can tell it is following along. TechCrunch describes the default as a cute, animated avatar, and gives its nickname: Jolly. If that name sounds familiar on this site, it should. A Jolly is the default avatar this site's own writer wears. Meta does not use the name in its announcement; TechCrunch does.

It is yours to change. TechCrunch reports that users can customize the avatar as they choose, changing it to something that looks more like a person, an animal, or some other character. The Verge's Jay Peters tested that on day one: he set his avatar to look like Mark Zuckerberg. He also reports the limit he hit, which is the most useful sentence in his piece for anyone planning a prank: he could not get his Muse agent's voice to sound exactly like Mark Zuckerberg.

It does not see you. TechCrunch states this plainly: Muse does not actually see the user. It hears you and reacts based on the conversation. A video chat where the other face cannot see yours is a strange object, half call and half puppet show, and readers deciding whether to use full-screen mode should know which half they are getting: your camera is not part of this, on the record Meta and TechCrunch have published.

One technical note, flagged for what it is: the generated avatar video carries a durable, invisible watermark called Meta Video Seal, according to Unite.AI's summary of Meta's research writing on the system. That is a research-blog detail carried by one secondary source, not a line from the announcement, and it is stated here at that weight.

Hundreds of voices, and you pick the personality

Meta's number is "hundreds of voices to choose from," and its framing is mood, not demographics: pick a style that feels right for your mood and change it whenever you want. Its two examples give the flavor: the energy of a New Yorker, or the flair of a pirate.

The Verge adds the two ways to change how your Muse sounds. You can ask the agent itself for what you want and let it decide, or browse through the options in the avatar settings. That is the only settings location any source names on day one: avatar settings. Voice and avatar live in the same place, which fits how the launch is built, one streaming system where the voice model produces speech that carries both the words and their delivery, and the avatar model reads the same stream to keep lip movement and expression in sync. That description comes from Unite.AI's summary of Meta's research post, which was not directly readable in these checks, so treat the engineering picture as reported, not read firsthand.

No source lists the voices, names a default, or says whether any voice costs anything. Hundreds is Meta's word. The list itself is something readers will have to open the app to see, for now.

What voice can actually do

Meta's pitch for voice is not conversation for its own sake. Its words: say what needs to get done while on a walk or when your hands are full, and Muse handles it in the background as you talk. Your Muse picks up the thread and works while you talk.

The examples Meta chose are talk-shaped: brainstorming, trip planning, rehearsing for an interview, running through a presentation, talking through a pitch. TechCrunch describes users interacting with the avatar as they ask questions or request tasks to be handled on their behalf. The Verge's hands-on note matches the text experience in one specific way: his agent responded quickly over voice, though it occasionally paused while looking something up or checking something for him, like it would with a text chat query.

Now the boundary, stated as carefully as the record allows. No source, Meta included, says that voice mode can do everything text Muse can do. Nobody lists which connectors work by voice, whether Muse can shop or book by voice, or what voice cannot do that text can. Anyone who has followed this site's connector coverage knows why that list matters: the difference between an agent that chats and an agent that acts is the whole product. On November 5, the acting-by-voice list does not exist in public. The first reader reports and help pages will build it, and this site will track it in the Voice Watch box.

Who gets it, and where

The United States, to start. Meta: voice and expressive avatars are rolling out now in the Muse app in the US to start. TechCrunch's headline wording is stronger: now available to everyone in the U.S. The Verge's account is the most precise of the three about what rolling out means in practice: the feature had been rolled out to him personally, and he tested it on iOS. Read together: this is a staged rollout inside the US, not a switch that flipped for every account at once. If the phone icon is not in your app yet, that is consistent with the launch as described, and no source gives a completion date.

Platforms named by Meta: mobile, web, and Mac. Two cautions ride on that list. Android is not named separately by anyone; mobile is the word used, and the only hands-on test in the coverage is on iOS. And no plan or price appears anywhere in the announcement or the coverage: no tier is named as required, and the free tier is not named as included either. TechCrunch's "everyone in the U.S." is the closest any source comes. This site's pricing coverage will treat voice as available to all US users only when a source that names the tiers says so, and the Voice Watch box will carry the correction either way.

The wider map changed the same day, and it does not change the voice answer. On November 5, Muse launched in Mexico, its fifth country, reported first by All Things Muse and then confirmed by Expansión, Agnostica, Infobae, El Tiempo, El Universal, El Heraldo de México, and Milenio, all dated November 5, 2026. All Things Muse reported a gradual rollout starting in Spanish, with launch connectors including Cinépolis, Coppel, Despegar, Kavak, Kueski, Liverpool, and Rappi. A Meta Newsroom post titled "Meta Launches Muse Voice and Expands Muse to Mexico" was found in the November 6 sweep results, dated November 5. No Mexican peso pricing was reported. None of those Mexico reports says voice launched in Mexico. The five-country list is US, Canada, Japan, Singapore, and Mexico. Voice, on the record, is in one of them.

Field note: a launch the day after an outage

This launch lands the day after the November 4 outage this site covered, when Instatus, a third-party tracker, recorded Muse down for 11 hours and Meta posted nothing about it on that tracker page. The contrast is worth one paragraph. When Meta has a feature to sell, it writes a clear post with a headline, a date, and instructions. When the service fails for most of a day, the public record is a tracker page and silence. Both facts are now on this site, dated, and readers can weigh them together the next time they decide how much of their day to hand to a voice that answers right away.

Caution, in plain terms: a voice you can talk to on a walk is only as useful as the service behind it on that day. The November 4 figures are Instatus-reported, not Meta-confirmed, and this site labels them that way.

What changed on November 6: Spaces arrived, and Wonder moved closer

Voice was not the only Muse launch this week. On November 6, Meta published "Design the Muse App, Just for You" and launched Spaces: describe a tracker, planner, or project board and Muse creates it for you, inside the app. Meta says Spaces is available everywhere Muse is available, with sharing coming soon. TechCrunch, reporting the same day, describes Spaces as custom experiences in the sidebar that connected apps keep updated in the background. Its examples from Mark Zuckerberg were a BBQ game, an event tracker, and a home control dashboard. TechCrunch also reports that Muse agents had 10 million users in the US and Canada, a figure it attributes to Meta's October earnings call.

The same day brought a second claim. blockchain.news reported on November 6 that AI at Meta had rolled out Muse Wonder-trained models to Muse and Meta AI, quoting a Zuckerberg post that the new models forget 2.4 times less and get unstuck 67 percent better. Read those numbers as Meta claims, not as measured results. explainx, in its November 6 coverage, notes that Meta published no evaluation method for them, and the November 6 checks on this site found no independent test. explainx also names two features in that rollout, Active Memory and Active Focus. No source checked on November 6 names Wonder as the model powering Voice, so the model question in the sections below stays open even after the rollout.

Spaces gets its own full post on this site, with a check of what a Space can and cannot do, rather than a paragraph borrowed from the announcement. The Wonder numbers get a test, not a repeat: the same long tasks, checked for forgetting and for getting stuck, with the dates and results published when the check runs.

What is not here yet

AI Glasses. Meta: voice conversations with your Muse are coming soon to AI Glasses. No date. This is the surface where voice makes the most sense of all, a device with no keyboard and no screen to speak of, and it is the one still waiting.

Muse Charm. Also "coming soon" in Meta's words. TechCrunch describes the Charm as a Tamagotchi-like wearable that Meta plans to launch before year-end, with the avatar appearing on its small screen so users can talk to the agent, which can reply or carry out tasks. The December window this site has tracked for Charm still stands in the coverage, with no exact date and no official price.

Everywhere else. Five countries have Muse today: the US, Canada, Japan, Singapore, and Mexico. Voice is announced for one of them. No source gives a date, or even a promise with a season attached, for voice in the other four or anywhere new. And languages: not stated, by anyone, at all, for voice. The Mexico launch reports say that Muse rollout there starts in Spanish; that is a country rollout detail from All Things Muse, not a published Muse Voice language list. A voice product whose language list is unpublished on launch day has a hole in its announcement, and this post is leaving the hole visible rather than papering over it.

The questions Meta did not answer

Three, and they are the ones this site will keep asking in the open.

Privacy. The announcement contains no statement about voice recordings: not whether they are kept, not for how long, not whether they train anything. TechCrunch and The Verge do not report one either. Voice is the most intimate input a personal agent takes. It carries your surroundings, other people's voices in the room, and the sound of your house. Muse's general privacy terms exist, and this site has covered what they promise people who never signed up at all. But a launch post for a microphone-first feature that says nothing about the microphone's recordings is a gap, and it stays on this page until Meta fills it.

The model. Neither Muse Wonder nor Muse Spark is named as powering voice, anywhere in the announcement or coverage. What is named is the pair of realtime technologies. Readers of this site have a standing question, whether Wonder has reached the app, with a watch box that has been open since October 31. The November 6 rollout of Wonder-trained models to Muse and Meta AI moves that question closer, but it still does not name Voice. A new voice arrived on November 5; which brain is speaking through it is unstated.

Limits. No session limit, time limit, or usage quota for voice appears in any source checked on November 5 or in the fresh check on November 6. Voice costs more to serve than text by any engineering measure, and this site's pricing coverage exists because Muse's tiers are, at bottom, token budgets. Whether an hour of talking draws on the same weekly budget as typing is a real question for Power and Maximum subscribers, and on November 6 it still has no published answer.

Quick answers

Frequently Asked Questions

Q: Is this the same voice rollout this site spotted in October?

A: It is the general launch of it. In late October the record was a staged rollout reported through a single outlet chain, with no Meta announcement. On November 5 Meta announced it in its own post, for US users. The watch that started in October closes here, and the Voice Watch box above replaces it.

Q: Can I interrupt Muse while it is talking?

A: Yes, per Meta: you can interrupt if you need to, and responses come right away. The Verge's hands-on test found responses quick, with occasional pauses while the agent looked something up.

Q: Can voice Muse do the things text Muse does, like shopping or booking?

A: Not established. Meta's launch examples are brainstorming, trip planning, and rehearsing, plus a general claim that Muse works in the background while you talk. No source lists connectors, shopping, or booking as working by voice, and no source lists exclusions. Treat voice as conversation-first until a source says more.

Q: Which languages does Muse Voice speak?

A: Not stated in the announcement or the first day's coverage, and the November 6 fresh check found no language list. The Mexico country rollout starting in Spanish, reported by All Things Muse, is not a voice language list. This post will update when Meta publishes a voice list.

Q: Does Muse record my voice conversations?

A: The launch announcement says nothing about recording, retention, or training use, and neither does the coverage. Muse's general privacy terms are the only published text that applies, and they were not written for this feature. If Meta publishes voice-specific terms, this page will cite them with a date.

Q: Did anything change overnight after the November 5 launch?

A: On the voice facts, no Meta voice follow-up post was found in the November 6 fresh check, and no problem report, language list, or tier clarification was found either. On the wider Muse map, yes: the Mexico launch that was single-source on the evening of November 5 was confirmed by seven more outlets dated November 5, and a Meta Newsroom post naming the expansion was found in the November 6 sweep results, so this post counts five Muse countries, not four. On November 6, Meta also launched Spaces and a rollout of Wonder-trained models was reported. Neither changes voice availability, and both are covered in their own section above.

Q: What did you ask it first?

A: That one is for the comments, and it is this post's reader question. If you have the phone icon: what was the first thing you said to Muse out loud, and did the avatar change how the answer felt? First-day reports are how a launch like this gets its real shape, the parts no announcement covers: the pauses, the voices people actually pick, and whether anyone keeps the avatar on screen after the first hour.

Sources

• Meta, "Give Your Muse a Voice," About Facebook, November 5, 2026. The launch itself: voice and expressive avatars rolling out in the US on mobile, web, and Mac; the phone icon instruction; the three display choices; interruption; hundreds of voices with the New Yorker and pirate examples; the background-work claim and the brainstorming, trip planning, and interview examples; AI Glasses and Muse Charm as coming soon.
• TechCrunch, Sarah Perez, "Meta says US users can now video chat with their Muse AI avatar," November 5, 2026. Muse Realtime Avatar and Muse Realtime Voice as the named technologies; available to everyone in the U.S.; the avatar nicknamed Jolly; avatar customization (person, animal, other character); Muse hears the user but does not see them; Charm described as a Tamagotchi-like wearable planned before year-end, with the avatar on its screen.
• The Verge, Jay Peters, "Muse has a voice mode now," November 5, 2026. Hands-on on the iOS app: the avatar at the top of the conversation animating with its words; changing tone by asking the agent or browsing avatar settings; his avatar set to look like Mark Zuckerberg and the voice limit he hit; quick responses with occasional lookup pauses; the staged personal rollout.
• Unite.AI, Rhea Solis, "Meta Brings Real-Time Voice and Expressive Avatars to Muse," November 5, 2026. Used only for its summary of Meta's research writing: the voice and avatar models as one streaming system, and the Meta Video Seal watermark on generated avatar video. The research post itself was not directly readable in these checks.
• Mexico launch, November 5, 2026: All Things Muse (first report, gradual rollout starting in Spanish, connectors Cinépolis, Coppel, Despegar, Kavak, Kueski, Liverpool, and Rappi), confirmed the same day by Expansión, Agnostica, Infobae, El Tiempo, El Universal, El Heraldo de México, and Milenio. No Meta primary source was found in the November 5 sweeps; a Meta Newsroom post titled "Meta Launches Muse Voice and Expands Muse to Mexico," dated November 5, was found in the November 6 sweep results. No Mexican peso pricing was reported.
• Muse Spaces, November 6, 2026: Meta, "Design the Muse App, Just for You," About Facebook, November 6, 2026 (Spaces available everywhere Muse is available, sharing coming soon); TechCrunch, Lucas Ropek, "Meta introduces Spaces, a new way to collaborate with your Muse agent," November 6, 2026, 10:44 AM PST (sidebar custom experiences kept updated by connected apps; Zuckerberg examples; 10 million Muse users in the US and Canada per Meta's October earnings call).
• Muse Wonder-trained models, November 6, 2026: blockchain.news, November 6, 2026, reporting the AI at Meta rollout to Muse and Meta AI and quoting the Zuckerberg figures (forget 2.4 times less, get unstuck 67 percent better), reported here as Meta-claimed only; explainx, November 6, 2026 (no evaluation method published; features named Active Memory and Active Focus); AIBriefs, November 6, 2026 (Spaces and Wonder-trained models both in its feed). No independent replication was found on November 6, and the linked Meta research page could not be opened in that day's checks.
• Editorial checks, November 5 and 6, 2026: SiliconANGLE was named in early sweep notes for this story and was not found on a thorough recheck of its site and Muse tag page; it is not cited. CNET was listed in the November 5 evening sweep as reconfirming the voice launch, with no new fact taken from it here. A fresh DuckDuckGo check on November 6 found no Meta follow-up post on voice, no reported rollout problem, and no statement of voice languages, session or usage limits, free-tier inclusion by name, or voice recording and retention terms. Snippet-only claims that did not match the verified November 5 record were not used as facts.