The interface between human and machine has gained a face — but it's one-way

Meta announced on November 5 that Muse Voice is now available to all users in the United States, enabling phone calls and real-time conversations with the animated avatar of its personal AI agent. The feature, first demonstrated at Meta Connect in September, runs on mobile, web, and Mac, and represents the company's next move in transforming Muse from a productivity tool into a companion with physical presence.

What Muse Voice actually is

What Meta calls a "videochat" is not, in fact, a video call in the traditional sense. When you initiate a conversation via the phone icon in the Muse app, the avatar that appears on screen — originally a cute white dog nicknamed "Jolly" — reacts to your speech with subtle facial expressions and gestures synchronized to the tone of your voice. The avatar speaks back, but it does not see you. The interaction is purely audio-driven: the system captures your speech, processes conversation tokens, and generates a visual response synchronized to the avatar.

The technology behind this is Muse Realtime Avatar, a Diffusion Transformer model that consumes the same voice token stream as Muse Realtime Voice to generate real-time visual performance. Instead of capturing a live camera feed from the user, the software renders facial expressions on screen based exclusively on what the user says. Meta describes the system as a unified streaming pipeline connecting conversational intelligence, voice synthesis, and visual embodiment — all running at about 870 milliseconds of latency, measured from the end of a user's utterance to the first synchronized frame.

The architecture behind the illusion

What makes Muse Voice technically notable is not the animation itself, but the scale and latency at which it runs. According to Meta's research blog, each generation produces eight frames corresponding to 320 ms of playback in 20 ms, or an effective 2.5 ms of model time per frame. The model streams portrait video at 448x768 pixels at 25 fps, and optimizations to the inference stack have increased serving capacity to 12 concurrent sessions on a single GB200 — an 8x jump compared to the two-step BF16 baseline.

But the engineering conceals a simple conceptual limitation: the avatar has no visual access to the world. It cannot see you waving, showing an object, or making a facial expression. It reacts to voice and conversational context, using gestures and expressions that match the tone. As Meta's own documentation describes it, this is a conversation with an animated character — not a video call with another person.

Customization and the anthropomorphization bet

Beyond adjusting voice tone — the app offers hundreds of options, from a "New Yorker" to a "pirate" — the user can completely redesign the avatar as a person, animal, or fictional character. CEO Mark Zuckerberg demonstrated the feature on Threads, and chief AI officer Alexandr Wang shared a video on X asking his avatar (a white dog) advice about the memes he posts. Meta's logic is clear: anthropomorphization increases emotional engagement. A friendly face encourages longer and more frequent interactions, and longer interactions generate more data to train better models.

This strategy is not new. Apple did it with Siri (which gained a 3D face in iOS 27), Google with the Gemis, and even OpenAI with ChatGPT avatars. What sets Muse apart is that it performs real-world actions — sending messages, managing tasks, making purchases — which raises the stakes of a deceptively expressive avatar hiding an agent with decision-making power.

The adult-only restriction and the trust challenge

The Muse app is strictly restricted to users over 18. Meta's justification is that the agent performs delegated actions (purchases, messaging, scheduling), which requires maturity to manage responsibly. There are no child-oriented safety controls — and this is not a technical omission but a deliberate product decision. If Muse is an agent that can spend money on your behalf and access your correspondence, Meta would rather not expose minors to a system that already has documented trust issues.

An article from October 20, 2026 revealed a vulnerability in Muse's sharing feature that exposed 6,800 conversations. Since then, Meta has added an invisible watermark via Meta Video Seal to all avatar-generated video — a safety measure against deepfakes that ironically also protects the platform itself: if someone tries to impersonate Muse, the watermark reveals the fraud.

What's next

Meta promises that Muse Voice will first arrive on its AI Glasses and the Muse Charm, a Tamagotchi-style wearable device expected to launch before year-end. The Charm will be a small portable device with a screen displaying the chosen avatar, enabling conversations without taking your phone out of your pocket. Meta says the same model stack powering the mobile and desktop version will run on the Charm.

The question is not whether the technology works — it does, with sub-second latency and coherent rendering — but whether users will actually want to talk to an avatar that cannot see them. Ultimately, Muse Voice is an experiment in digital presence: how much face, voice, and gesture is enough to make someone feel accompanied by a machine?

Sources: TechCrunch, The Verge, Meta AI Research

✓ Independent sources cross-checked and verified before publishing