Skip to main content

Smart Turn Detection

The Smart Turn Detection system is an advanced feature that uses machine learning to intelligently determine when a speaker has finished their turn in a conversation. This system enhances the accuracy of endpoint detection beyond traditional time-based thresholds.

Configuration Parameters​

enable_turn_detection​

  • Type: Boolean
  • Default: false
  • Description: Enables or disables the smart turn detection feature.

When enabled:

  • The system uses an ML model to predict turn completion, in addition to the standard timeout-based method.

When disabled:

  • The system relies only on turn_detection_timeout (in seconds).
  • Example: if turn_detection_timeout = 0.6, a 0.6-second pause is treated as the end of the turn.

turn_detection_timeout​

  • Type: Float (seconds)
  • Default: 0.6
  • Description: The timeout threshold for end-of-speech detection. This value is stored as eos_threshold_seconds and used as a fallback when smart turn detection is disabled or when the ML model confidence is low.

2. Server Options​

options = {
    "enable_turn_detection": True,
    "turn_detection_timeout": 0.6,
}

3. Environment Variables​

  • EOS_THRESHOLD_SECONDS: Default timeout value (default: 0.6)

Benefits​

  1. Improved Accuracy: ML-based detection is more accurate than time-based thresholds alone
  2. Context Awareness: The model understands speech patterns and context
  3. Reduced False Positives: Better at distinguishing between natural pauses and actual turn completion
  4. Adaptive: Works across different speakers, languages, and speaking styles
  5. Fallback Safety: Traditional timeout still works as a safety net

Prediction Interpretation​

  • Probability > 0.95: Very high confidence - immediate EOS trigger

Best Practices​

  1. Enable for Conversational AI: Most beneficial for interactive applications
  2. Tune Timeout: Adjust turn_detection_timeout based on your use case
  3. Test with Real Data: Validate performance with your specific audio characteristics
  4. Fallback Strategy: Always keep timeout as a safety net