Meta began rolling out spoken conversations and animated avatars in its Muse personal agent app on November 5, 2026, two months after Muse launched and six weeks after the company previewed the feature at its Connect conference. According to Meta's Newsroom post, titled "Give Your Muse a Voice," the feature works in the Muse app on mobile, on the web and on the Mac, offers "hundreds of voices," and lets Muse carry out tasks in the background while a person is still talking. The rollout covers the US only "to start," although Muse itself is available in the US and Canada. Voice conversations are "coming soon" to Meta's AI glasses and to Muse Charm, a pocket device that Meta has not yet priced or dated.

In Short

Meta's Muse app, which can do tasks like shopping or sending emails for you, can now hold a spoken conversation and show a cartoon-like character that reacts while you talk. It matters because a growing number of people may soon ask an agent for things out loud while walking or cooking, rather than typing into a search box or scrolling a feed. For now it only works in the US, and Meta's announcement does not say what happens to the voice recordings or whether they have anything to do with ads.

What Meta switched on

The mechanics are simple. According to Meta, people start a voice session with Muse by tapping a phone icon: "just tap the phone icon in the top-right corner to start talking." From there, the conversation runs in real time. Meta credits two systems it calls Muse Realtime Voice and Muse Realtime Avatar, which together, according to the post, allow people to "Hear responses right away. You can even interrupt if you need to."

Interruption matters because many earlier voice assistants waited for a full spoken turn before answering. Meta does not give a response time in the announcement itself. It does link to a research post, discussed below, that does.

The second feature is that Muse keeps working while the conversation continues. One of the four takeaways at the top of the post says people can "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." That is the difference between Muse and a conventional voice assistant: Muse is an agentic AI product that books, buys and sends things on a person's behalf, and the voice layer now sits on top of those actions.

Meta's examples are domestic rather than commercial. The post lists brainstorming, trip planning and rehearsing for job interviews, presentations and pitches. A 30-second demonstration clip on the page shows a user asking, "Can we role play my interview, then email the recruiter to confirm the date?" Muse answers: "Let's do it! I'll play the hiring manager." A second, 59-second clip shows a cooking exchange in which the user asks for recipe ideas for an "Italian queso" and Muse suggests "Queso e Pepe," which it describes as "A Texas spin" on the dish.

Three ways to see the avatar

The avatar is optional. According to Meta, users choose between "voice only, the avatar alongside your conversation, or full screen when you want to talk face to face." The post says the avatar "uses subtle expressions and gestures that match the tone of the conversation" and "reacts while you talk, so you can tell it's following along."

In both demonstration clips, the avatar is a furry, cream-coloured character rather than a human face: it wears a baseball cap with an Italian flag in one clip and a cowboy hat in the other. The voice post does not say whether users can design their own avatar or pick a human-looking one. Meta published a separate post a day later, on November 6, titled "Design the Muse App, Just for You," which introduced custom tools called spaces.

Hundreds of voices

The voice itself is configurable. "With hundreds of voices to choose from, you can pick a style that feels right for your mood and change it whenever you want," according to Meta. The post's examples are "the energy of a New Yorker or the flair of a pirate." It does not publish the list, say how many languages or accents are covered, or say whether any voices are licensed from real people.

From Connect preview to rollout

The feature has been public for six weeks. Meta's recap of Connect 2026, the company's annual hardware and developer event, which opened on September 23, listed a voice mode for Muse and argued that talking to an agent is different from voice chat with an assistant. That recap gave no launch date for voice. It also said Meta would bring Muse to its AI glasses "in the coming months" and that it would have more to share about Muse Charm later in the year.

On Meta's third-quarter earnings call on October 21, Mark Zuckerberg named Muse Voice among the models and products the company had released in recent months, and said Muse would come to Meta Glasses starting the following month, as PPC Land reported from the call. He described Muse Charm as a device that would go from concept to people's pockets in less than a year, without a launch date.

The November 5 post is less precise. It says: "Voice conversations with your Muse are coming soon to AI Glasses and Muse Charm." It does not name which glasses models, and it does not say whether the avatar will appear on any device beyond phones and computers.

The technology behind the avatar

The words "Muse Realtime Avatar" in the announcement link to a Meta AI Research post titled "Bringing Your Muse to Life," dated September 23, 2026, the day of Connect. That post carries the technical figures the announcement leaves out.

According to the research post, Muse Realtime Voice produces a stream of speech tokens. An audio decoder turns that stream into sound, and the avatar model reads the same stream, which is how Meta keeps lip movement and facial expression in step with the voice. The avatar model is an audio-driven diffusion transformer that takes three inputs: the speech tokens, a reference image that can be a portrait, an illustration or an object, and a rolling window of recently generated video. It produces video in short chunks, one after another.

Speed came from distillation. The original "teacher" model, according to Meta, needed 40 diffusion steps with three guidance passes each, or 120 model evaluations per chunk. The production "student" model needs two, a reduction Meta puts at 60 times, and Meta says the smaller model closely preserves the teacher's quality.

The headline figure is latency: about 870 milliseconds from the end of the user's turn to the first byte of the synchronized voice and video reply, according to the research post. The other output figures are just as specific:

  • Video is generated at 448 by 768 pixels in portrait orientation, at 25 frames per second.
  • On a single Nvidia GB200, each generation step yields eight frames, or 320 milliseconds of playback, in 20 milliseconds of compute.
  • After four-bit quantization and other serving work, one GB200 can run 12 concurrent real-time video sessions, eight times the baseline Meta started from.

Twelve sessions per chip is a useful number for scale. At that rate, 1,000 people talking to a full avatar at the same moment would occupy about 84 GB200s, by PPC Land's calculation. Meta does not say how many users choose the avatar over voice only, or how many run sessions at the same time.

Meta also published comparisons with two commercial avatar products. In 2-to-3-minute live conversations, raters in Meta's tests preferred Muse over Runway Characters 78% to 22% and over HeyGen LiveAvatar 88% to 12%, according to the research post. These are Meta's own evaluations of its own product, with no outside review, and on one measure, mannerisms, the post says the gap with Runway was not statistically distinguishable from parity.

Two further details in the research post bear on how the output can be used. Generated video carries Meta Video Seal, the company's watermarking system. And the research post states that Muse is for users aged 18 and over, a condition the November 5 announcement does not mention.

What the announcement leaves out

Several gaps stand out once the post is read against Muse's history.

Canada. Muse was available in the US and Canada by September 29, when Meta announced Muse for small business, and Zuckerberg gave a combined US and Canada user count on October 21. The voice post limits the rollout to the US without saying when Canadian users will get it. On the October call, Zuckerberg said Muse had not launched globally because it is expensive in compute to run, and a feature that can generate 25 frames of video per second for each user in avatar mode adds to that bill.

WhatsApp. When Muse launched on September 8, PPC Land reported that it was reachable on iOS, Android, muse.ai and directly inside WhatsApp. The voice post refers only to "the Muse app on mobile, web and on Mac." WhatsApp is not mentioned. The Mac was not among the launch surfaces in September either.

Price. Muse has a free tier and paid plans for heavier use. The post does not say whether voice, the avatar or particular voices are limited to either.

Approvals. Meta's September 29 post on Muse for small business said nothing publishes, sends or spends without the user's approval, a commitment PPC Land reported on October 11. The voice demo includes a request to email a recruiter mid-conversation. The post does not explain how approvals work when the user is speaking rather than tapping.

Recordings. The post does not say whether voice audio is stored, for how long, whether transcripts are kept, or whether recordings are used to train Meta's models.

Voice, Muse and advertising

That last gap matters more for Meta than for most companies, because of what it has said before.

At Muse's launch, Meta said conversations and the data inside each person's Muse Secure VM, the cloud machine that runs the agent, are not shared with its ad systems, and that people can opt out of having their interactions used for model training, as PPC Land set out on September 9. The voice post neither repeats that commitment nor says whether spoken conversations fall under it.

The question has a precedent at Meta. In October 2025, the company said that from December 16, 2025, people's interactions with its AI features would be used to personalize content and ads, and it explicitly included voice chats alongside text exchanges, as PPC Land reported at the time. That policy covered Meta AI. Muse was set up with a different promise.

The direction of travel is not fixed, either. On the October 21 call, Zuckerberg said people who want Muse to improve the personalization of their content and ads across Meta would be able to connect those experiences, without explaining how that connection would work or what data would cross, according to PPC Land's account of the call. Earlier in October, an MBI Deep Dives analysis argued that Muse may never need ads of its own yet could still shape Instagram ads, because the agent's browsing of merchant sites can feed merchant targeting.

Meta's stated revenue model for Muse is different again. On the same call, Zuckerberg said the product should be free for most people and that Meta would collect fees, paid by businesses, on transactions Muse helps complete. Shopping is already Muse's most common use, he said, with more than a quarter of people using Muse engaged in it and an average purchase above $200.

Two months of Muse

The voice rollout lands on a product that has grown quickly and drawn friction just as quickly.

Muse launched in the US on September 8, 2026. Within 12 days, Amazon had cut it off from buying on Amazon.com, complaining among other things that the agent did not identify itself when browsing. Sensor Tower estimated that Muse was downloaded more than 5 million times in the US and Canada within 22 days, and HUMAN found that Muse accounted for 72% of the AI agent requests it observed in its early data.

By October 20, Sensor Tower put Muse at 10 million US downloads in 40 days and 2.5 million US daily active users, with Meta directing 39% of its October advertising spend to promoting the app. The same day, Meta halted Muse share links after 6,800 turned up in public DNS records. A day later, Zuckerberg told analysts more than 10 million people in the US and Canada had set up Muse agents, and that about 25% of them use it on any given day.

Voice changes how people reach the agent. Meta's own framing, "Voice means Muse goes with you," points at more occasions to use it: on a walk, while cooking, or with hands full.

Rivals already talk

Meta is not first to put voice on an AI app. Anthropic began a beta of voice conversations in Claude's mobile apps in May 2025, initially in English. Google gave Gemini for Mac voice control over any open app in July 2026. The animated, interruptible avatar is the less common part, and Meta's own research post benchmarks it against specialist avatar companies rather than against other assistants.

The more human an assistant sounds, the closer it comes to rules written for companion chatbot products. New York Governor Kathy Hochul on October 28 signed a law that bars operators from offering minors 12 features it classes as unsafe, including outputs suggesting the system is human, and its definition explicitly covers systems that interact through writing or sound, as PPC Land reported. It takes effect on January 1, 2027. California's SB 243, signed in October 2025, already requires a notice where a reasonable person could be misled into thinking a chatbot is human. Meta's research post says Muse is for adults, and its demo avatars are plainly not human, which limits the overlap. Whether Muse counts as a companion product under either law has not been tested.

Why this matters for marketers

Voice changes the shape of the request. A typed query to an agent tends to be short and specific. A spoken one, by Meta's own description, starts with "the half-baked idea," and the agent "picks up the thread and works while you talk." For brands whose products get chosen inside Muse, that means recommendations formed during a loose conversation, with no results page and, by default, no visible citation trail. Profound reported in October that four of Muse's six typical shopping picks came from two browsed websites rather than from Meta's own catalog, and that Muse does not name the sources it visits by default.

Voice also moves Muse off the screen. If conversations reach glasses and Muse Charm as promised, more agent activity could happen with no visual interface at all, leaving no place for a brand to appear through display or search ads. Meta's stated answer is transaction fees paid by businesses. No fee level, start date or merchant terms have been published.

And voice adds a data question that Meta has not answered. Meta announced in October 2025 that voice interactions with Meta AI would inform ad personalization from December 16, 2025. Muse began with a commitment that its conversations stay out of ad systems. Whether spoken Muse conversations sit under that commitment, the older Meta AI policy, or something new is the most important detail missing from the November 5 post.

Timeline

Summary

Who: Meta, through its Muse personal agent app. The feature affects Muse users in the US, businesses whose products Muse recommends or buys, and advertisers on Meta's platforms.

What: Meta added real-time voice conversations and an animated avatar to the Muse app. Users can interrupt, choose from hundreds of voices, and pick voice only, an avatar alongside the chat, or a full-screen avatar, while Muse carries out tasks in the background. A linked Meta research post puts response latency at about 870 milliseconds and capacity at 12 concurrent avatar sessions per Nvidia GB200.

When: Announced and rolling out from November 5, 2026, after a preview at Connect on September 23, 2026. Voice on AI glasses and Muse Charm is "coming soon," with no date.

Where: The Muse app on mobile, web and Mac, in the US only to start. Muse is also available in Canada, where voice has not been switched on. WhatsApp is not mentioned.

Why: Meta is making Muse usable away from the screen, ahead of glasses and a dedicated pocket device. The announcement does not say how voice recordings are handled or whether they are kept out of Meta's ad systems, as Muse conversations have been since launch.