Meta says one Nvidia GB200 can support 12 concurrent avatar sessions for Muse voice chats, while the first video frame and accompanying audio arrive about 870 milliseconds after a user finishes speaking.
Two models split the job: Muse Realtime Voice packages a reply's words and vocal tone, and Muse Realtime Avatar turns that output into synchronized video, Meta says. At 25 frames per second, the avatar generates portrait video measuring 448 pixels wide and 768 pixels tall, Meta says. For visual consistency across a conversation, it creates short video segments and consults frames from its earlier output.
Meta says a reduced model derived from a larger one now performs two computations instead of 120. The original required 40 steps of three calculations each; Meta said the smaller model's image quality stays close to the larger model's.
Meta said a GB200 takes 20 milliseconds per pass to create eight frames, which make up 0.32 seconds of video. The company listed four-bit quantization among its optimizations and separately reported that optimized serving capacity was eight times baseline.
On Thursday, Nov. 5, U.S. users could start Muse voice chats in the app's mobile, web and Mac versions, Meta said. The interface allows mid-response interruptions and an audio-only setting; the avatar can appear next to the messages or occupy the full display. Meta says its app has hundreds of voice choices.
Meta said evaluators tested Muse against Runway Characters and HeyGen LiveAvatar in two- to three-minute chats with the same character. The company reported that evaluators selected Muse in 78% of matchups with Runway and 88% with HeyGen. Meta said the demeanor ratings for Muse and Runway showed no statistically significant difference.
Meta identified its AI glasses and Muse Charm as future destinations for voice chats but supplied no release dates.
