Voice & Turn-Taking Settings

These settings control how aggressively the agent detects user speech and decides when to respond.

Interruptions

• allow_interruptions
  • Allows the user to interrupt the agent mid-speech.
• min_interruption_duration
  • Minimum duration of detected user speech to count as an interruption.

Defaults:

• allow_interruptions=True
• min_interruption_duration=0.08

Turn-taking / endpointing

These parameters control Voice Activity Detection (VAD) and determine when the agent decides the user has finished speaking:

• min_silence_duration - How long (in seconds) the user must be silent before the system considers they might be done speaking.
• activation_threshold - Confidence level (0.0-1.0) required to detect speech. Higher values reduce false positives from background noise.
• prefix_padding_duration - Amount of audio (in seconds) to capture before detected speech starts. Helps catch the beginning of words.
• min_endpointing_delay - Minimum time (in seconds) to wait before ending the user's turn.
• max_endpointing_delay - Maximum time (in seconds) to wait before forcefully ending the turn.

Defaults:

• min_silence_duration=2.0
• activation_threshold=0.4
• prefix_padding_duration=0.5
• min_endpointing_delay=0.45
• max_endpointing_delay=3.0

These defaults are optimized to handle natural conversation pauses and prevent the agent from cutting off users who are thinking or formulating their thoughts.

Practical guidance

• If the agent responds too quickly: Increase min_silence_duration or max_endpointing_delay.
• If the agent cuts off users mid-thought: Increase min_endpointing_delay and max_endpointing_delay.
• If users experience silence detection issues: Check if the stream is closing prematurely by increasing max_endpointing_delay to 4.0-5.0 seconds.
• If the agent interrupts users too easily: Reduce allow_interruptions or increase min_interruption_duration.
• If speech is missed at the beginning: Increase prefix_padding_duration.
• For noisy environments: Increase activation_threshold to 0.65-0.7 to reduce false speech detection.
• For users who speak slowly or thoughtfully: Increase both min_silence_duration (to 1.0-1.2) and max_endpointing_delay (to 4.0-5.0).