Meta rolls out real-time voice and expressive avatars in its Muse app
On November 5, 2026, Meta began a US rollout of live voice chat and animated avatars for its Muse agent across mobile, web and Mac, according to Unite.AI. Muse, the personal AI agent Meta introduced on September 8, 2026, works through the Muse app or WhatsApp, and the voice layer was first previewed at Connect 2026 that September. To start a call, users tap a phone icon in the app's top-right corner, then choose voice-only mode, an avatar beside the chat, or a full-screen avatar. Hundreds of voices are available, and users can change them at any time.
Muse Realtime Voice gives immediate spoken replies and lets people interrupt mid-answer, and Muse keeps working on requests while a user walks or cooks. The avatar shows facial expressions and gestures that respond as the user speaks. Meta's stated use case is hands-free planning, such as brainstorming while walking or when hands are full. Meta has not said whether voice or avatars require a paid plan.
Unite.AI, citing a Meta AI Research technical post, reports that distillation cut avatar generation from 120 guided teacher evaluations per chunk to two unguided student evaluations. The same work reports eightfold higher serving capacity than a two-step BF16 baseline and 12 concurrent real-time video sessions on a single GB200. Generated video carries an invisible Meta Video Seal watermark, and Meta warns that its examples show model capability rather than every avatar in the app.
Why it matters
For founders, the rollout shows voice and an animated face becoming part of the default consumer agent package rather than a side experiment. The update is a product change rather than a new model, so answer quality still depends on the model underneath. Teams should expect users to judge assistants on how natural the conversation feels, not only on what the assistant can answer.
For operators and investors, the practical work sits in latency, interruption handling and turn-taking, since those decide whether people keep talking. Meta's figures point to heavy compute demands for real-time avatar video, which makes serving efficiency a margin question as well as a technical one. Buyers and funders will likely ask how voice and avatar usage is priced, and Meta has not yet said whether these features need a paid plan.
Key points
- 1Rollout began November 5, 2026 in the US on mobile, web and Mac.
- 2Users start a call by tapping the phone icon in the top-right corner of the app.
- 3Users can choose voice-only mode, a side avatar, or a full-screen avatar, with hundreds of voices.
- 4Distillation cut avatar generation from 120 guided teacher evaluations per chunk to two student evaluations.
- 5Meta reports eightfold serving capacity over a two-step BF16 baseline and 12 concurrent sessions on one GB200.
