What is a production voice avatar?
A browser joins a WebRTC room. A GPU agent drives lip-synced frames from speech. TTS audio is cleaned so markdown bullets do not create acoustic hangs. RAG supplies the words before the mouth moves. If any of those layers is missing, you have a video, not an avatar you can talk to.
Proof: Voice Avatar Platform
The CIC Academy / MuseTalk LiveKit deployment issues JWTs from a Flash serverless endpoint, lists avatars from network volume, and reports CUDA health. A persistent MuseTalk agent runs loading, greeting, idle, listening, thinking, and speaking states with eased blending. The browser sees the live feed, transcript, connection status, and retrieval confidence. Clients never talk to the GPU pod directly.
Avatar vs phone receptionist
Use an avatar when presence and visual attention matter—tutoring, product walkthroughs. Use a phone agent (MrBot, Peoplix) when the job is booking or an HRIS write. Combining both without a clear job creates a expensive novelty.
Related work and reading
Frequently asked questions
How does lip-sync stay realtime?
MuseTalk runs on GPU (for example RTX-class via RunPod) and streams frames into the LiveKit room while TTS plays. Onset/offset blending keeps transitions from popping.
Can the avatar use private documents?
Yes. RAG retrieves first; the avatar speaks grounded content suitable for tutoring and demos.
Is this only a research demo?
The stack is production-shaped: persistent agent states, Flash deploy hooks, network volumes for avatars, and an operator session UI.