Meta rolls out real-time voice for Muse

research.meta.ai10 min read

Meta AI Research has introduced Muse Realtime Avatar, embodiment technology that converts speech into expressive, interactive avatars. The system uses an audio-driven Diffusion Transformer conditioned on speech tokens, reference media, and recent video latents to generate synchronized streaming video. Meta reports that distillation techniques reduce inference steps while custom serving optimizations enable low latency and high concurrency on NVIDIA hardware. Open article →

More on this topic