Skip to main content
Realistic Avatars give your voice agent a photorealistic face. Send your agent’s voice as it’s generated; your users see a real-looking person say it, with accurate lip sync, natural expressions and subtle head motion, streamed to your page in real time.
  • Any voice: OpenAI, ElevenLabs, Azure, Fish Audio, Cartesia, or any other TTS or speech-to-speech model. The avatar follows the audio, so no text or timings are needed.
  • Real time: speech appears on the avatar’s lips moments after you send it. Interruptions are one call.
  • Voice and video in sync, delivered together over WebRTC. No media servers to run.
  • Drop-in for LiveKit: the avatar can join your LiveKit room as a participant instead.
  • A few lines of code on your server and your page. The SDK handles streaming, playback and reconnection.
Realistic Avatars include no LLM and no text-to-speech: you bring your own voice agent. Want a complete agent with no code? See Voice Agents.

​
How it works

  1. Your server starts a session with your API key and an avatar ID, then pushes your agent’s audio to it as it streams.
  2. Talking Avatar renders the avatar speaking that audio, with the voice and lips in sync.
  3. Your page plays the video and voice, or the avatar joins your LiveKit room with them as tracks.
Your API key never leaves your server. Your page only plays the stream.

​
Two ways to show the avatar

​
The SDK

Both modes use @interviewflowai/talking-avatar. Your API key must stay on your server, so the SDK has two parts:

​
Not on Node.js?

Use the realtime avatar API directly: one WebSocket, from any language. It does everything the SDK does. There’s a complete Python example, and your page can still use LiveAvatarView, or plain WebRTC.

​
Video

All avatars stream at 832 × 468 (16:9), 25 frames per second. LiveAvatarView fits the video inside its container without cropping.

​
Pricing

$0.02 per minute, billed per second from your prepaid balance. A session is billed from the moment the avatar is live until you close it. Each session can run for up to 3 hours, and you can run as many as you need. See Pricing.

Quickstart

A complete server and page, step by step.

Set up with a coding agent

A ready prompt for Claude Code, Codex or Cursor.