Muse gains real-time voice and an expressive avatar

Meta rolls out Muse with voice conversations and expressive avatars in the United States on mobile, web, and Mac. The agent can respond, interrupt, and continue tasks in the background during the exchange.

Since November 5, 2026, Meta has been deploying real-time voice conversations and expressive avatars in Muse, its personal AI agent, in the United States on mobile, web, and Mac. The service, already showcased at Connect 2026 in September, adds a spoken and visual interface to an assistant that previously mainly acted via text. The challenge goes beyond a gadget: Meta wants to make Muse a tool where you speak, correct, and where some tasks continue while you chat.

When speaking replaces the neat prompt

With Muse Realtime Voice, users can start a conversation by tapping the phone icon at the top right, then speak freely without crafting a perfect query. Meta cites very concrete cases: bringing out an incomplete idea, planning a trip from a city or a few preferences, rehearsing a job interview, a presentation, or a pitch. The model responds immediately, accepts interruptions, and keeps control over tasks assigned in the background. Convenient, especially when cooking, walking, or busy hands; the kind of moment when the keyboard is less handy.

An avatar that listens, speaks, and follows the tone

The other novelty is Muse Realtime Avatar, which gives a face to the exchange. Users can stay in voice-only mode, display the avatar next to the conversation, or switch to full screen for a face-to-face version. Meta says it offers hundreds of voices, with varied styles, including a “New Yorker” profile and a “pirate” one, and the selection remains changeable at any time. The avatar reacts while listening and speaking, with expressions and gestures meant to match the tone of the discussion.

What Meta puts behind the showcase

Technically, Meta presents Muse Realtime Voice and Muse Realtime Avatar as a single streaming system. The voice model produces speech tokens that carry both content and how it is said, then an audio decoder transforms these tokens into voice while the avatar model generates the synchronized image. This architecture must keep speech, lip movement, and expressions aligned without breaking the conversation’s rhythm.

The associated research paper also describes an avatar conditioned by this token stream, a reference image, and a rolling window of recent video. Practically, the system produces short video blocks, then reinjects the visual context produced in the previous block to maintain gestures and appearance. Meta says it reduced generation from 120 teacher-guided evaluations to two unguided student evaluations, a 60-fold reduction while maintaining quality close to the reference model.

A strategy preparing glasses and pocket format

Meta explains that voice conversations will also arrive “soon” on AI Glasses and Muse Charm, two formats designed for hands-free use and pocket interaction. The first is meant to extend Muse into connected glasses; the second into a compact device dedicated to conversation. The most interesting question remains: if voice, interruption, and avatar really hold up in real time, Muse will no longer be just an agent that responds, but an assistant you can almost interrupt like a busy colleague.

The rollout starts, however, with an American version on classic screens, and that is where real adoption will be played out: less on demonstration than on daily fluidity, reliability of delegated tasks, and the ability not to give the impression of having three interfaces for the price of one.

Key points

  • November 5, 2026: launch in the United States on mobile, web, and Mac.
  • Connect 2026, on September 23, had shown the mode before release.
  • Meta announces hundreds of voices, modifiable at any time.
  • The model accepts interruption without closing the exchange.
  • Muse is reserved for users aged 18 and over.
  • Meta mentions upcoming arrival on AI Glasses and Muse Charm.

By the numbers

  • 870 milliseconds — latency measured between the end of a user turn and the first synchronized response, according to Meta AI Research.
  • 448 x 768 pixels — output format of the real-time video portrait, according to Meta AI Research.
  • 25 frames per second — claimed output rate for the avatar.
  • 60 times — announced compute reduction after distillation compared to the teacher model.
  • 12 sessions — concurrent real-time video sessions supported on a GB200, according to Meta.

To read