Skip to content
Back to showcase

Omni Turn Detector

A low-latency model that ends the agent's turn sooner, predicting the boundary from the shape of the audio instead of waiting for silence.

Role
Solo, model and inference
Timeline
2026
Stack
Rust · ONNX
Status
Live in production
Omni Turn Detector

Fig. 01 · InterfaceRust · ONNX · WebRTC

The problem

Waiting for silence is waiting too long. A voice agent that only speaks once the user has clearly stopped will always feel a beat behind the conversation.

What it does

Predicts the end-of-turn boundary from the shape of the audio, rather than waiting for a fixed pause to elapse.

The fastest turn is the one you saw coming.

PreviousDocument PortalCase study 01 / 03
Next project

Inference Server

Low latency inference service for hosting AI models

  • Rust
  • gRPC
  • Tokio
  • Protobuf