Meta rolls out voice and expressive avatars in the Muse app

What happened?

Meta said November 5 that it is rolling out voice and expressive avatars in the Muse app on mobile, web and Mac, starting in the United States. Users can tap the phone icon to begin talking, choose voice-only or avatar views and interrupt the conversation. Meta says voice conversations with Muse are coming soon to AI Glasses and Muse Charm; the announcement does not say those surfaces are available now.

The product announcement builds on a September launch that introduced Muse in the US on iOS, Android and muse.ai, with basic use free and subscription plans for heavier use. The new voice layer is described as a way to think aloud, plan, rehearse and continue a conversation while walking or doing another task. Meta says the avatar uses expressions and gestures that match the conversation and offers hundreds of voices, including stylized options.

Meta’s product language also defines the interaction model: the avatar reacts while the user is speaking, and the user can choose voice only, a companion avatar or full-screen conversation. That makes the release more than a talking-head demo, but it also means the interface is designed to sustain attention. The relevant question for users is not merely whether the animation looks natural; it is whether the system makes its generated nature, permissions and limits clear for everyone during a persuasive conversation.

Meta AI Research’s September technical post is broader than the product announcement. It describes Muse Realtime Avatar as a streaming system that conditions on reference media and can animate photographic portraits, full-body illustrations, animals and everyday objects. The post reports 448x768 portrait video at 25 frames per second, about 870 milliseconds from the end of a user turn to the first byte of synchronized voice and video, and 12 concurrent sessions on a GB200 after optimization. In a two- to three-minute rater comparison with Runway Characters and HeyGen LiveAvatar, raters preferred Muse overall, while the mannerism comparison with Runway was not statistically distinguishable from parity.

Source: Meta ↗ · Source: Meta AI Research ↗ · Source: Meta ↗ · Source: Meta AI Research ↗

Why does it matter?

The most important product boundary is the line Meta itself draws between research and rollout. A technical demonstration can show what the system can generate under a controlled setup; it does not establish that every reference image, animal animation, latency number or concurrency condition is available on a consumer device. Meta’s research page explicitly says the examples do not all reflect avatars available in the Muse app. The November announcement is the stronger source for current access, and it only promises a US-first rollout on specified surfaces.

The latency figure is also a systems measurement, not a guarantee of a user’s perceived response time. Meta measures the first byte of synchronized voice and video after a turn, while a person experiences network conditions, audio capture, buffering, rendering and interruptions. The 12-session GB200 figure is a serving-capacity result on Meta infrastructure. Those details make the research credible as an engineering report without turning it into an independent consumer benchmark.

For users and teams, Muse is a shift from a text or voice assistant toward a more present character interface. That can make rehearsal and brainstorming feel more natural, but it increases the importance of disclosure and control: people need to know when an expressive response is generated, what media is being animated and which permissions are active. Meta says Muse uses Video Seal for an invisible durable watermark and describes connected-app approval, audit trails and controls for training use. Those are vendor-stated safeguards, not an independent privacy or safety audit, so the practical question is whether the controls remain legible when the avatar is persuasive and fast.

Source: Meta ↗ · Source: Meta AI Research ↗ · Source: Meta ↗ · Source: Meta AI Research ↗