Conversation behavior and call controls
Conversation behavior is produced by several layers working together. The system prompt defines policy and dialogue flow. Speech settings determine when a caller's turn is complete and whether an interruption stops playback. Lifecycle controls handle prolonged silence and maximum duration. Functions perform actions such as transferring a live call.
Do not try to control all four layers with prompt wording. A sentence such as “wait five seconds” does not configure a timer, and writing a function marker does not attach or execute a function.
In this topic
- Configure turn taking and interruptions — distinguish response timing, barge-in, backchannels, and architecture-specific controls.
- Configure silence and call limits — set the silence prompt, inactivity hang-up, and hard duration cap.
- Design greetings and conversation flow — configure the first spoken turn and write concise, recoverable dialogue.
- Transfer and end calls safely — define human handoff, supported transfer modes, closing behavior, and current voicemail boundaries.
- Test and troubleshoot call behavior — evaluate timing, interruption, silence, transfers, and call completion end to end.
Control map
| Desired behavior | Configure it in |
|---|---|
| What the agent should say or ask | System prompt |
| Exact first spoken message for an inbound call | Custom Greeting |
| How quickly a Modular recognizer commits a caller turn | Response Speed / STT patience |
| How many caller words stop Modular playback | Words to Interrupt |
| Short acknowledgements that should not stop Modular playback | Filler Words to Ignore |
| Responsive, Balanced, or Patient Realtime turn detection | Voice Delivery turn taking |
| Ask whether the caller is still present | Silence Timeout |
| End after continued inactivity | Hangup Timeout |
| Enforce a maximum total call length | Limit Call Duration |
| Move the live call elsewhere | Attached Call Transfer function |
Architecture boundary
Modular and Realtime calls share lifecycle controls, but they do not share the same interruption mechanism.
- Modular uses streaming transcription, an interruption word threshold, filler/backchannel filtering, and an STT patience profile.
- Realtime uses provider-native voice activity and interruption events. Its Responsive, Balanced, or Patient turn-taking control belongs under Voice Delivery. Modular word thresholds and filler lists are inactive.
Capability notices in the editor indicate whether a control is native, pipeline-managed, guided, or inactive for the selected architecture.
Recommended configuration sequence
- Write the role, outcomes, boundaries, and action rules.
- Configure the greeting.
- Choose the architecture and speech settings.
- Tune turn taking with fixed test scripts.
- Set silence and call-duration limits.
- Attach transfer or message-taking functions where required.
- Test the draft in the browser and on the phone network.
- Publish with a useful version note and monitor real calls.