livekit / agents Realtime

Voice agents that answer while you are still talking.

livekit/agents is a framework for building realtime voice AI agents. The whole design problem is the gap between a person finishing a sentence and an agent starting one.

Build an agent github.com/livekit/agents
One turn, end to endIllustrative pipeline
Speech inaudio arrives over the room connection
Turn detectiondecide the person actually stopped
Transcribespeech becomes text the model can read
Reasonthe agent picks what to say and do
Synthesizetext becomes audio, streamed early
Speech outthe reply plays back in the room

Stage order is the pipeline shape, drawn illustratively. No timing numbers are claimed here.

Realtime

Latency is the interface

A voice agent is judged on the pause before it speaks, which makes every stage in the pipeline a design decision.

Framework

You bring the agent

The framework owns transport, turn handling, and the audio path, so your code can be about what the agent should say.

Voice

Built for spoken turns

Interruptions, overlaps, and half-finished sentences are normal input here rather than error cases.