Copy page

Guides

Conversation Behavior

The conversation section of the agent config controls how calls sound and feel — when the agent speaks and listens, how interruptions are handled, and what background sounds are played.

Turn Detection

Turn detection determines when the agent decides the caller has finished speaking.

Code
{
  "conversation": {
    "turn_detection": {
      "mode": "semantic_vad",
      "min_delay": 0.5,
      "max_delay": 6.0,
      "endpointing_mode": "dynamic"
    }
  }
}

Mode	How it works	Best for
vad	Silence-based detection — waits for a pause in speech	Fast, low-latency
semantic_vad	ML model predicts whether the utterance is complete	More natural, handles partial sentences better

Parameters:

Field	Default	Description
min_delay	0.5s	Minimum silence before triggering (prevents cutting off fast speakers)
max_delay	6.0s	Maximum wait before the agent responds anyway
endpointing_mode	dynamic	fixed uses a constant threshold; dynamic adapts to conversation pace

First Message

Configure what the agent says when the call connects:

Code
{
  "conversation": {
    "first_message": {
      "enabled": true,
      "text": "Hello! Thank you for calling Acme Corp. How can I help you today?",
      "delay_ms": 400,
      "allow_interruptions": false
    }
  }
}

Field	Default	Description
enabled	true	Whether the agent speaks first
text	"Hello! How can I help you today?"	Opening message. Supports {{variables}}.
delay_ms	400	Pause before speaking (ms). Gives caller time to settle.
allow_interruptions	false	Whether the caller can cut off the greeting

Interruption Handling

Controls what happens when the caller speaks while the agent is talking:

Code
{
  "conversation": {
    "interruption": {
      "enabled": true,
      "mode": "vad",
      "min_duration": 0.5,
      "min_words": 0,
      "false_interruption_timeout": 2.0,
      "resume_false_interruption": true,
      "discard_audio_if_uninterruptible": true
    }
  }
}

Field	Default	Description
enabled	true	Allow callers to interrupt the agent
min_duration	0.5s	Minimum speech duration to count as an interruption
min_words	0	Minimum words before interruption triggers
false_interruption_timeout	2.0s	If caller goes silent within this window, resume the interrupted response
resume_false_interruption	true	Resume if the "interruption" was just background noise
discard_audio_if_uninterruptible	true	Drop buffered audio when agent is mid-sentence and can't be interrupted

Silence Timeout

Hang up if the caller goes silent:

Code
{
  "conversation": {
    "timeout": {
      "enabled": true,
      "trigger_seconds": 15.0,
      "warning_count": 2,
      "warning_messages": ["Are you still there?", "Hello?"],
      "final_message": "I'll end the call now. Feel free to call back anytime. Goodbye!"
    }
  }
}

The agent sends up to warning_count messages before ending the call.

Call Duration Limit

Prevent calls from running indefinitely:

Code
{
  "conversation": {
    "call_duration": {
      "enabled": true,
      "max_seconds": 1800,
      "warning_seconds_before": 60,
      "warning_message": "We're approaching our time limit. Let me wrap up quickly.",
      "exceeded_message": "We've reached the maximum call time. Thank you for calling!"
    }
  }
}

Maximum supported: 7200 seconds (2 hours). Minimum: 60 seconds.

Backchannel Responses

Natural listening cues ("mm-hmm", "I see") inserted during long caller utterances:

Code
{
  "conversation": {
    "backchannel": {
      "enabled": true,
      "interval_seconds": 5.0,
      "responses_short": ["mm-hmm", "uh-huh", "right"],
      "responses_long": ["I understand", "Go on", "I see"]
    }
  }
}

Background Audio

Play ambient sounds to make the call feel more natural:

Code
{
  "conversation": {
    "background_audio": {
      "ambient": {
        "enabled": true,
        "source": "office_ambience",
        "volume": 0.3
      },
      "thinking": {
        "enabled": true,
        "source": "keyboard_typing",
        "volume": 0.5
      }
    }
  }
}

Available presets:

Preset	Description
office_ambience	Quiet office background
city_ambience	Urban street sounds
forest_ambience	Natural outdoor sounds
crowded_room	Busy indoor environment
keyboard_typing	Keyboard sounds during LLM processing
keyboard_typing_2	Alternate keyboard variant
hold_music	Gentle hold music

Set source to a URL to use a custom audio file.

Ambient plays continuously throughout the call. Thinking plays while the LLM is generating a response — helps mask latency.

Preemptive Generation

Start generating the next response before the caller finishes speaking:

Code
{
  "conversation": {
    "preemptive_generation": false
  }
}

Disabled by default. Enable for lower perceived latency, but may result in responses that don't fully address the complete utterance.

TTS Text Transforms

Filter content before sending to the voice engine:

Code
{
  "conversation": {
    "tts_text_transforms": ["filter_markdown", "filter_emoji"]
  }
}

filter_markdown removes **bold**, # headers, and similar markup. filter_emoji removes emoji characters.

Both are enabled by default to prevent LLM-generated formatting from being read aloud.

Related

• Agents — full agent configuration
• Configuring Voice & Speech — TTS and STT provider settings
• Post-Call Analysis — analytics configuration