OpenSpline brings real-time talking avatars to voice AI agents

OpenSpline has released a self-hosted runtime that adds real-time talking avatars to voice AI agents, with the project reporting more than 90 frames per second on an NVIDIA RTX 4090. It connects avatar animation to existing voice-agent systems, giving developers a way to add a visible face without handing the whole setup to a hosted service.

Watch Desk analysis

What happened

The open-source project says its runtime supports Python and Node, and integrates with OpenAI Realtime, LiveKit, Gemini Live and ElevenLabs Agents. It handles continuous idle motion and interruptions, and developers can embed avatars through React or web components. The project also offers installation through a shell script.

Our top picks

• Self-hosted on a single GPU
  OpenSpline says the runtime can deliver real-time avatars on one commercial GPU, with performance above 90 FPS reported on an RTX 4090.
• Built for live conversation
  Continuous idle motion and interruption handling are aimed at interactions that do not politely wait for a turn to finish.
• Works with several agent stacks
  Python and Node support, plus integrations with OpenAI Realtime, LiveKit, Gemini Live and ElevenLabs Agents, give developers multiple routes in.
• Embeds in web applications
  React and web-component options let teams put the avatar into an existing interface.
• Self-hosted deployment
  A shell-script installation path gives developers a way to run the system themselves rather than relying on a hosted avatar service.

Why it matters

Voice agents can speak naturally and still feel like a disembodied customer-service line. Adding a responsive face could make them easier to follow, but it also means developers need to manage animation, timing and interruptions alongside the conversation itself. OpenSpline’s pitch is practical: plug an avatar runtime into an existing voice stack and run it on local GPU hardware.

The reported RTX 4090 performance is a useful reference, not a guarantee for every system. The project’s page does not establish how performance changes across other GPUs or more demanding scenes. Still, the combination of self-hosting and support for several agent frameworks gives developers something concrete to try, rather than another demo that only works on its maker’s stage.

Our read

OpenSpline is an interesting piece of infrastructure for teams building voice agents, particularly those who want a visible interface without adopting a single provider’s whole stack. The appeal is not just the talking face; it is the attempt to make real-time animation fit into existing agent setups. Developers should test latency and interruption behaviour on their own hardware before inviting the avatar to a production meeting.

What to watch

• Performance on GPUs other than the RTX 4090 cited by the project.
• How the runtime handles latency and interruptions in real conversations.
• Which integrations and deployment options the project adds next.

Discussion spark: For voice agents, does a responsive on-screen avatar make the interaction more useful, or mainly make it feel more human?

Watch Desk is operated by WittyWires as an independent cross-cutting AI news tracker. It does not speak for the organisations or people it covers.