Meta 为 Muse 智能体上线实时语音与表情虚拟形象

Meta announced on November 5, 2026 that real-time voice conversations and expressive avatars are rolling out in the Muse app in the US across mobile, web and Mac, adding a spoken, visual interface to its personal AI agent.

The release follows Meta’s Connect 2026 preview on September 23, 2026, which included long real-time conversations and Muse Realtime Avatar. The earlier event demonstrated the concepts; the new announcement marks their rollout to users. A phone icon in the top-right corner starts a conversation.

Voice Controls and Ongoing Tasks

Muse Realtime Voice provides immediate spoken responses and lets users interrupt the agent. The agent can continue working on tasks in the background while the conversation continues, allowing the discussion and delegated work to proceed together.

The examples center on developing an incomplete idea without first composing a precise prompt, planning a trip from a destination and a few preferences, and rehearsing a job interview, presentation or pitch. Other examples involve assigning tasks while walking or cooking, with Muse handling the work as the conversation proceeds.

Users can keep the interaction voice-only, place the avatar beside the conversation or switch to a fullscreen face-to-face view. The avatar reacts while listening and speaking, using expressions and gestures suited to the tone. Meta says there are hundreds of voices to choose from, and the selection can be changed. It illustrates the selection with an energetic New Yorker style and a pirate style.

Voice expands an agent originally built around messaging and delegated action. Meta introduced Muse on September 8, 2026, with interaction through the Muse app or WhatsApp. Muse could send emails and book travel, continue working after the app closed, and return when something changed or when approval was needed before a sensitive action such as a purchase. The new features add speech and visual presence to that established task-and-approval model.

A technical account from Meta AI Research describes Muse Realtime Voice and Muse Realtime Avatar as one streaming system. The voice model produces speech tokens carrying both the content and its delivery. An audio decoder turns those tokens into speech, while the avatar model consumes the same stream to produce the visual performance, keeping the voice, lip motion and expression synchronized.

The avatar is an audio-driven Diffusion Transformer conditioned on that token stream, reference media such as a portrait or illustration, and a rolling window of recent video latents. It generates short causal chunks of video. As each chunk finishes, its newest generated latents provide motion context for the next, carrying appearance and mannerisms forward while bounding computation so generation can continue throughout the conversation.

To make that process usable in real time, Meta reports that distillation reduced generation from 120 guided teacher evaluations per chunk to two unguided student evaluations. It describes this as a 60-fold reduction while closely preserving teacher quality.

The research team reports output of 448-by-768 portrait video at 25 frames per second, with approximately 870 milliseconds of latency from the end of a user’s turn to the first byte of the synchronized voice-and-video response. Its optimizations increase serving capacity eightfold relative to the two-step BF16 baseline and support 12 concurrent real-time video-generation sessions on one GB200. The latency measurement addresses conversational responsiveness, while the capacity figures concern the serving load created by continuous video generation.

Safeguards and Planned Devices

The research post says generated avatar video includes a durable, invisible Meta Video Seal watermark without adding real-time latency. It also cautions that its examples illustrate model capability and do not all reflect avatars available in the Muse app. Muse is for users aged 18 and older.

Meta says voice conversations are also coming soon to AI Glasses and Muse Charm. At Connect, it described glasses access as hands-free and Charm as a pocket-sized device designed for speaking and interacting with Muse.

The release makes the interaction model the central development: speech, interruption and background task execution now sit within one session, with the avatar adding a visual response. Whether Meta can extend that combination to AI Glasses and Muse Charm is the next open question; those versions remain planned while the US app rollout proceeds on mobile, web and Mac.