Meta’s assistant speaks and comes to life on screen

Voice and expressive avatars are coming to Muse in the U.S. to interact with Meta’s assistant on mobile, the web, and Mac.

The voice arrives in Muse with a character that accompanies the conversation on screen. Meta’s assistant can respond orally, be interrupted, and react visually while its user speaks. This new conversation mode brings together two features that users can separate: talking without a keyboard and giving a visible presence to their artificial interlocutor.

After a presentation at Connect 2026, the rollout is beginning in the United States on mobile, on the web, and in the Mac app. No schedule is indicated for France. The service is intended for people aged 18 and over. Access is via a phone icon at the top right of the interface. Voice conversations are also announced for AI Glasses and Muse Charm, with no launch date specified. An idea can take shape through the exchange

The use case highlighted consists of developing a thought by speaking, even when the initial request remains vague. To prepare a trip, for example, the user can mention a destination, specify the type of stay desired, then add their wishes during the discussion. The request is built progressively, rather than requiring carefully prepared text before the first exchange.

The ability to interrupt the assistant is important in this process. A response can be redirected without waiting for it to finish, to correct information or change direction. According to Meta, the assistant also carries out requested tasks in the background during the conversation. The press release describes use while walking or preparing a meal, but does not detail the concrete actions performed in these situations.

Preparing to speak is another use presented. A job interview, a presentation, or a professional pitch can be rehearsed out loud, with an answer to react to immediately. The value here lies in the format of the exercise: hearing one’s interlocutor and responding to them forces one to work differently than by rereading a succession of messages. A character on screen, if the user wants it

The visual component remains optional. Three displays are planned: voice only, an avatar shown next to the conversation, or a full-screen character. The user therefore chooses the place occupied by this animated presence, depending on whether they want to keep their eyes on the interface or simply listen.

The movements are not limited to the lips. The character adopts expressions and gestures associated with the tone of the conversation; it also reacts while the user is speaking. These visual cues constitute a staging of listening, without in themselves establishing what the assistant understands from the exchange. The choice to display them remains separate from the choice to use voice.

Customization also comes through a catalog of several hundred voices, according to Meta. The examples range from a New York-style way of speaking to a pirate voice, and the choice can be changed at any time. The setting thus concerns both the style of the interlocutor and the way a request is addressed to it. Voice and animation produced together

Behind this interface, the work on Muse Realtime Avatar describes a joint production of sound and image. A single stream represents the content of speech and the way it is spoken. It is used to generate the voice and then the corresponding animation, in order to synchronize lip movements and expressions. The video is produced in small successive segments, while preserving recent visual context to maintain the character’s appearance and gestures.

The researchers announce video at 25 frames per second and a latency of about 870 milliseconds. This measurement runs from the end of the user’s turn to the receipt of the first fragment of response combining voice and video. It describes their technical system, without guaranteeing this delay for every conversation in the app. The research demos also present characters based on illustrations, animals, or objects; Meta specifies that these examples do not all correspond to the avatars actually offered in Muse.