Meta opens US access to Muse voice calls with expressive, animated avatars

Meta began offering real-time voice and video conversations with its Muse agent to users across the United States on 5 November, pairing a speech model with an animated avatar that modulates its expressions and gestures as it speaks. The feature, Muse Voice, extends the consumer AI agent Meta released in September and moves the assistant away from the text box: TechCrunch reports Meta's argument that anthropomorphising the agent in this way could make consumers feel more connected to it, a point that puts the product in the same competitive field as voice-capable agents from OpenAI and Anthropic. Meta says the conversations work on mobile, web and Mac.

Two models do the work. Muse Realtime Voice generates speech tokens carrying both what is said and how it is delivered; Muse Realtime Avatar then synchronises voice, mouth shapes and expression through an audio decoder and an avatar model. The video is produced by a Diffusion Transformer-based model that, according to AI Times, Meta accelerated with a distillation technique cutting the teacher model's diffusion process from 120 inference steps to two, yielding 448×768 portrait video at 25 frames per second and roughly 870 milliseconds of latency after the user stops speaking. AI Times also reports that Meta tuned inference with Nvidia to run 12 video sessions on a single GB200 GPU, eight times the serving capacity of its baseline model. The system processes only the user's audio and does not access the camera, even during a video call, ANSA reports; Meta further embeds an invisible Video Seal watermark in generated avatar footage to trace its provenance.

The avatar defaults to Jolly, a small dog, and can be reshaped into a person, an animal or a fantasy character. Users choose from hundreds of voices — ITmedia cites "New Yorker" and "pirate" styles — select between audio-only, in-conversation and full-screen displays, and can interrupt the agent mid-sentence. Access is restricted to over-18s.

Reports on the current footprint do not fully align. ITmedia said Muse reached Japan on 23 October, while ANSA listed four active markets: the United States, Canada, Japan and Singapore. In a separate report, Infobae said Meta had launched Muse in Mexico, the first Latin American market to receive it, with local connectors including Rappi, Liverpool and Cinépolis. The Verge's reviewer found voice responses quick, with occasional pauses while the agent looked something up, and the animation "a little uncanny" once the avatar was styled after Mark Zuckerberg. ITmedia framed the distinction from ChatGPT and Claude as the ability to hand the agent work by voice while an expressive avatar supplies something closer to a face-to-face exchange, noting that the rivals' voices come mainly from preset options whereas Muse offers hundreds.

Meta says the features will come "soon" to its AI glasses and to Muse Charm, a Tamagotchi-like wearable due to ship before the end of the year, whose small screen will show the avatar as users talk to the agent and assign it tasks. That shipping deadline is the next checkpoint.