Voice Autopilot turn detection and interruptions
Voice turn detection and playback handling control when Voice Autopilot listens, responds, accepts interruptions and completes announcements.
What it is
VAD Mode controls how Voice Autopilot detects when a caller speaks and finishes a turn. Playback handling also controls interruptions and waits for important announcements to be heard. This helps keep the spoken conversation in step with the telephone call.
Who uses it and why
Administrators choose turn detection and announcement wording. Callers can interrupt ordinary replies, while operators receive calls after the configured hand-off sequence.
Where to find it
Open Settings › AI › Features and edit a Voice Autopilot feature. Use the Settings role with AI access or the Developer role. The voice options include VAD Mode, Initial Greeting Message and the final-announcement fields.
How it works
Automatic mode uses the speech service to detect speech and pauses. Semantic mode uses the spoken words to help identify the end of a turn. Manual mode uses caller audio levels and transcription checks. Available options depend on the configured voice option list.
When Initial Greeting Message is set, the agent speaks that exact opening once. The opening is protected against interruption. If a caller makes a substantive request during it, INSIDE defers one reply until the greeting and caller turn finish. A greeting-only utterance such as "hello" does not prompt a repeated greeting.
During ordinary replies, caller interruption clears queued telephone audio. INSIDE adjusts the model's conversation history to the part played to the caller. Audio discarded from the queue does not count as heard content.
For configured final announcements, INSIDE stops competing replies and waits for telephone playback before transferring, ending the call or starting voicemail. These waits are bounded so a missing audio connection cannot hold the action indefinitely.
Configure it
Choose VAD Mode to suit your call testing. Automatic is the default. If the available alternatives do not appear, ask Powerfront to confirm the voice options available in your deployment.
Set Initial Greeting Message when you need fixed opening words. Leave it blank to use the agent's instructions.
Set Human Transfer Message, Request Callback Message, Record Voicemail Message and Call End (Goodbye) Message when those actions need exact wording. Blank values retain the existing instruction-led behaviour. Max Duration Transfer Message defaults to blank and uses this built-in announcement: "Thanks so much for your patience. I'm now connecting you with a member of our team who will continue to help you. Please stay on the line."
Settings
| Setting | Where | Applies to | Default | Who can change it | What it does |
|---|---|---|---|---|---|
| VAD Mode | Settings › AI › Features | website | Automatic (OpenAI Server VAD) | Settings role, Developer role | Selects how Voice Autopilot detects caller speech and the end of a turn. |
| Initial Greeting Message | Settings › AI › Features | website | Not set | Settings role, Developer role | Supplies the exact opening greeting; blank lets the agent use its instructions. |
| Max Duration Transfer Message | Settings › AI › Features | website | Not set: uses the built-in announcement, Thanks so much for your patience. I'm now connecting you with a member of our team who will continue to help you. Please stay on the line. | Settings role, Developer role | Sets the announcement before human queue hand-off when maximum duration is reached; blank uses the built-in announcement. |
| Human Transfer Message | Settings › AI › Features | website | Not set | Settings role, Developer role | Supplies an exact announcement for a successful human transfer; blank retains instruction-led behaviour. |
| Request Callback Message | Settings › AI › Features | website | Not set | Settings role, Developer role | Supplies an exact confirmation after callback registration and before ending the call. |
| Record Voicemail Message | Settings › AI › Features | website | Not set | Settings role, Developer role | Supplies an exact announcement after voicemail preparation and before recording starts. |
| Call End (Goodbye) Message | Settings › AI › Features | website | Not set | Settings role, Developer role | Supplies an exact closing line before the agent hangs up; blank lets the agent use its instructions. |
Permissions
| Permission | Lets you |
|---|---|
| AI | With the Settings role, allows voice feature configuration; the Developer role can also access Features. |
Limits and good to know
- Opening greetings and final announcements are protected; ordinary replies can be interrupted.
- Generating audio is not the same as playing it. A final action can wait while already-generated audio finishes.
- Final-announcement start waits are up to 12 seconds by default, or 4 seconds for queue-transfer announcements. The playback wait is up to 15 seconds. These are wait bounds, not response-time promises.
- Missing audio, a disconnected call or a timeout can prevent an announcement from completing.
- Transcription availability affects how caller speech during the greeting is classified. Fallback handling can continue when a transcript does not arrive.
Depends on and works with
View all connections as a list
This feature uses
- Voice Autopilot call handlingConfigured by
The voice feature supplies greetings, final announcements and permitted actions.
- Voice Autopilot connection setupRequires
Turn detection and playback synchronisation operate within a realtime Voice Autopilot session.
Used by
- Voice Autopilot call handlingConfigured by
Speech turn-taking and playback affect interruptions and spoken handover messages.
- Voice Autopilot connection setupConfigured by
Turn detection, greetings and playback handling govern spoken interaction.
FAQ and troubleshooting
Why did the caller fail to interrupt the opening greeting?
The opening greeting is deliberately protected. Ordinary replies use the normal interruption handling.
Why did the agent answer a request made during the greeting later?
INSIDE waits for the greeting and caller turn to finish, then responds once to a substantive overlapping request.
Why is there a pause before transfer or hang-up?
INSIDE can be waiting for the configured announcement to start and finish playing. Those waits have time limits.
Why was the final announcement cut short?
Check call connectivity and the configured message. Announcements are best effort and can time out. Ask Powerfront to inspect the affected call if it recurs.