Skip to main content
In a natural conversation, people often talk over each other. “Barge-in” is the ability for a user to interrupt the AI while it is speaking. Implementing this correctly is crucial for making an AI feel “real.”

​
How it Works

When using the Lokutor VoiceAgentClient, the system is designed to handle this loop:
  1. User Speaks: Your application continues to stream audio input.
  2. VAD Detection: Our Voice Activity Detection (VAD) identifies that the user has started speaking.
  3. Interrupt Signal: The server immediately stops generating audio for the current response.
  4. Buffer Clearing: Your client application must clear its local playback buffer to stop the audio immediately.

​
Implementation Guide

The Lokutor SDKs handle the interruption signal automatically at the protocol level. However, you must ensure your audio playback queue is cleared. 
client.onStatus((status) => {
  if (status === 'interrupted') {
    // 1. Stop your local speaker/audio player
    audioPlayer.stop();
    // 2. Clear any pending audio chunks
    audioBuffer.clear();
    console.log("Agent was interrupted by user.");
  }
});

​
Manual WebSocket Implementation

If you are communicating directly via WebSocket, you should listen for an interruption event: 
{
  "type": "event",
  "event": "agent_interrupted",
  "reason": "user_speech_detected"
}
Upon receiving this:
  1. Drop all chunks: Discard any binary chunks currently in your network buffer.
  2. Kill Playback: Stop the current audio playback stream immediately.

​
Best Practices

  • Latency is Key: The faster you detect the user’s voice and kill the agent’s playback, the more natural it feels. We recommend a local VAD (like Silero or WebRTC VAD) in addition to our server-side detection for the fastest possible response.
  • Graceful Fade: instead of a hard cut, a very fast (50ms) linear volume fade-out can make the interruption feel less jarring to the user.