Meta brings voice conversations and expressive avatars to Muse in the US
Meta has added real-time voice conversations and expressive avatars to its Muse app, the personal AI agent the company launched in the United States in September. The feature was announced on the Meta Newsroom on November 5, 2026, after a preview at Connect 2026, and it is rolling out first to users in the US on mobile, on the web and on Mac.
The new mode lets users talk to Muse out loud, hear replies immediately and interrupt the agent mid-answer, while an animated avatar reacts to the conversation with expressions and gestures matched to its tone. Meta says users can choose from hundreds of voices in different styles and switch between them at any time. The avatar can be shown next to the conversation, expanded to full screen or turned off in favor of a voice-only call.
Two models power the experience: Muse Realtime Voice, which produces the spoken replies, and Muse Realtime Avatar, a video generation model that animates the avatar from the same stream of speech data. According to the technical description Meta Superintelligence Labs published in September, the avatar streams portrait video at 448x768 pixels and 25 frames per second, with roughly 870 milliseconds between the end of a user's turn and the first byte of the synchronized voice and video reply.
Voice conversations are also listed as coming to Meta's AI glasses and to Muse Charm, the pocket device Meta showed at Connect, but the company has not given a date for either.
··········
MUSE VOICE CONVERSATIONS: WHO GETS THEM AND HOW TO START ONE.
The rollout follows the same footprint as the Muse app itself. Muse launched in the US on iOS, Android and the muse.ai website, and Meta's November 5 announcement adds Mac to the list of places where voice conversations work. A call starts from the phone icon in the top-right corner of the Muse interface, and the display mode can be changed during the conversation.
Meta has not said that voice or avatars require a paid plan. The company describes Muse as free for most of what people need, with subscription plans for heavier use, and the voice announcement does not attach any price, usage cap or plan restriction to the new mode. Supported languages, daily limits and the size of the avatar catalog in the app are also absent from the announcement. Meta's research write-up states that Muse is for users aged 18 and over, and that some of the avatar examples it showed demonstrate model capability rather than avatars that are available in the app.
........
Item | What Meta has stated |
Announcement date | November 5, 2026 (Meta Newsroom) |
Regions | United States, to start |
Platforms | Mobile, web and Mac |
How to start | Phone icon, top-right corner of Muse |
Display modes | Voice only, avatar beside the chat, full screen |
Voices | Hundreds, changeable at any time |
Interruptions | Supported |
AI glasses and Muse Charm | Coming soon, no date |
Price or plan limits | Not disclosed |
Languages | Not disclosed |
........
For people already using Muse as an agent, the more practical change is that the voice channel does not pause the agent's work. Meta says Muse can keep handling tasks in the background while the user is still talking, a behavior it first described at Connect as the main difference from the turn-by-turn voice modes of conventional assistants. Tasks that touch the outside world keep the approval steps of the regular app, where Muse asks for confirmation before actions such as sending an email or completing a purchase, and a separate Sentinel agent has to approve any outbound internet access. Readers who want the background on those controls can find it in our report on Muse payments, the secure VM and the Sentinel agent.
··········
HOW MUSE REALTIME AVATAR TURNS SPEECH INTO VIDEO.
Meta's research team explained the avatar model in a September 23 blog post that the company now links to the consumer launch. Muse Realtime Voice emits a stream of discrete speech tokens that encode both the words and the way they are delivered. An audio decoder converts those tokens into sound, and Muse Realtime Avatar reads the same token stream to generate the face and body movement, which is how Meta keeps lip motion, expression and voice in step with one another.
The avatar model is an audio-driven Diffusion Transformer. Besides the speech tokens, it is conditioned on reference media, which is the image the avatar is built from, and on a rolling window of recently generated video latents. Video is produced in short causal chunks, and the newest latents of each chunk become the motion context for the next one, so the amount of computation per step stays bounded however long the conversation runs. Meta says the reference media can be a portrait, a full-body illustration, an animal or an everyday object.
Real-time generation required heavy distillation. Meta first trained a bidirectional teacher model that uses 40 diffusion steps with three-way classifier-free guidance, which adds up to 120 model evaluations for every chunk of video. A causal student with a fixed-length KV cache was then trained with self-forcing and distribution matching distillation, and it runs two unguided evaluations per chunk. That is a 60x reduction in model evaluations. In Meta's own preference tests the student and the teacher came out close to even, with the reported values falling in a 45% to 55% band.
........
Muse Realtime Avatar specification | Value reported by Meta |
Output video | 448x768 portrait, 25 fps |
Latency, end of user turn to first reply byte | About 870 ms |
Frames per generation step | 8 frames (320 ms of playback) |
Time per generation step on one GB200 | 20 ms |
Effective model time per frame | 2.5 ms |
Concurrent real-time sessions per GB200 | 12 |
Serving capacity vs two-step BF16 baseline | 8x |
Model evaluations per chunk, teacher vs student | 120 vs 2 |
........
The serving figures come from Meta's measurements on NVIDIA's GB200 hardware. Each generation step produces eight frames, or 320 milliseconds of playback, in 20 milliseconds of compute, so the model generates video about 16 times faster than it is played back on a single accelerator. Meta attributes the headroom to a set of engineering choices that includes persistent KV caches, memory-aware positional encodings, cache-aware routing, latency-aware dynamic batching, four-bit quantization-aware training, fused kernels and NVIDIA CUDA Graph capture, and it says it worked with NVIDIA on the optimizations. The result is 12 concurrent real-time video sessions on one GB200, eight times the capacity of the two-step BF16 baseline Meta started from.
··········
MUSE REALTIME AVATAR EVALUATION AGAINST RUNWAY AND HEYGEN.
Meta compared its avatar with two commercial live-avatar products, Runway Characters and HeyGen LiveAvatar, using each product's native live-call experience. The avatars were given matched identities, and testers held conversations of two to three minutes before rating the results. In overall preference, Muse Realtime Avatar was chosen in 78% of comparisons against Runway Characters and in 88% against HeyGen LiveAvatar.
Raters preferred Meta's avatar on every dimension measured, according to the company, although the comparison on mannerisms against Runway Characters was statistically indistinguishable from parity. These are vendor-run evaluations. Meta designed the protocol, chose the identities and reported the results, and no independent replication has been published, so the percentages describe preference under Meta's test conditions rather than a general quality ranking.
··········
MUSE AVATARS COMPARED WITH GEMINI LIVE AVATAR.
The closest offering from another large lab is Google's Gemini 3.8 Live with Live Avatar, announced on September 24, 2026. The two products target different users. Google positions Live Avatar as an enterprise feature inside Gemini Enterprise, with preset avatars, custom avatars generated from a reference image for allowlisted enterprise customers, and API documentation for developers. Meta's avatars sit inside a consumer app and are aimed at individual users in the US.
........
Feature | Muse voice and avatars (Meta) | Gemini 3.8 Live with Live Avatar (Google) |
Announced | November 5, 2026 (consumer launch) | September 24, 2026 |
Audience | Consumers in the US | Gemini Enterprise customers |
Avatar source | Built-in avatars in the Muse app | Preset library; custom avatars for allowlisted enterprises |
Languages | Not disclosed | 97 languages, speech to speech |
Background work during a call | Yes, Muse continues tasks | Asynchronous tool calling |
Published latency | About 870 ms to first reply | Not disclosed |
Watermark | Meta Video Seal, invisible | SynthID in audio and video |
Price | Not disclosed | Not disclosed |
........
The comparison is not like-for-like. Meta has published latency, frame rate and serving numbers but not language coverage for the consumer feature, while Google has published language coverage but no latency figure. Neither company has attached a price to its avatar feature in the announcements, and the two products serve different buyers.
··········
SAFETY CONTROLS AND WATERMARKING IN MUSE AVATARS.
Video produced by Muse Realtime Avatar carries an invisible watermark from Meta Video Seal, which the company says adds no latency to the live experience. Meta also says it enforces strict safety requirements across the avatar experience to reduce the risk of misuse, without listing those requirements in detail, and that it will keep strengthening the protections. The research post does not explain whether users can create an avatar from a photo of a real person in the consumer app, and the Newsroom announcement does not describe custom avatar creation at all. Users should treat the avatars offered inside Muse as the supported options until Meta documents anything further.
··········
OTHER MUSE APP CHANGES ANNOUNCED THIS WEEK.
The voice launch came alongside two other Muse updates. On November 6, Meta introduced Spaces, which lets users describe an idea and have Muse build a custom experience inside the app, such as a trip planner, a school dashboard or a small-business order tracker. Spaces can be pinned to the sidebar or main tabs, hidden or scheduled to appear at certain times, and connected apps can keep them updated in the background. Meta says Spaces are available everywhere Muse is available and that it is exploring ways to share them with other people.
Meta also extended Active Memory and Active Focus to all Meta AI and Muse users, a change covered in our report on Active Memory and Active Focus for Meta AI and Muse. Readers comparing the new mode with Meta's earlier voice assistant can refer to our guide to Meta AI voice mode and how it compares.
··········
MUSE VOICE AND AVATARS: CONFIRMED FACTS AND OPEN QUESTIONS.
The launch itself is documented by Meta. Voice conversations with expressive avatars are live in the Muse app for US users on mobile, web and Mac, with hundreds of selectable voices, interruption support, three display modes and background task handling during a call. The technical figures for the avatar model, including the 870-millisecond response latency, 25 fps portrait output, 12 sessions per GB200 and the 60x cut in model evaluations, come from Meta's own research post and measurements, as do the preference results against Runway Characters and HeyGen LiveAvatar.
Several points remain undisclosed. Meta has not said which languages the voice mode supports, whether heavy use counts against the limits of Muse subscription plans, when the feature will reach countries outside the US, or when it will arrive on AI glasses and Muse Charm. The company has also not detailed how it prevents avatars from imitating real people beyond the Video Seal watermark and a general commitment to safety requirements. Until those details are published, the feature is best understood as a US consumer launch of a real-time avatar system whose performance claims rest on Meta's internal testing.
··········
FOLLOW US FOR MORE.
·····
DATA STUDIOS
·····
datastudios.org
