Speech and conversation behaviour
Choose Pipeline or Realtime speech and configure turn-taking, interruptions and limits.
Start with a working provider stack. Change one setting at a time and compare calls with representative caller audio.
Choose a speech mode
| Mode | Use it when | Requirements and limits |
|---|---|---|
| Pipeline | You need separate transcription, model and voice choices, validated answers or structured intake. | Compatible providers for each stage. |
| Realtime | You want direct OpenAI audio-to-audio conversation. | A compatible GPT Realtime model and OpenAI realtime voice; validated-document answers and structured flows require Pipeline. |
Realtime retains tool approvals and authorized retrieval. Its audio choices differ from ordinary TTS voices. A model supporting text chat does not automatically support realtime audio.
Tune turns and interruptions
In Transcriber, select turn completion, silence timing, noise filtering and supported provider hints. Short silence can cut off hesitant callers; long silence makes replies feel slow. Test pauses within names, phone numbers and Kiswahili phrases.
Configure interruption behaviour and response timing in Advanced. Set duration, inactivity, response-count and stage-deadline limits suitable for the task. An interruption should stop speech cleanly and let the next turn proceed.
Pronunciation entries map written words to spoken aliases in Pipeline TTS without changing caller transcripts. Realtime treats pronunciation guidance as model instructions. Test county and business names with the chosen voice.
Fallbacks
Configure ordered transcription fallbacks and supported model/voice recovery. Every fallback needs a working connection in your organisation. Primary-provider hints do not automatically transfer to a different provider.
After switching from a failed transcriber, the assistant may ask the caller to repeat. A fallback does not safely retry an external action with an unknown outcome.
Measure the result
Use Generate preview for voice settings and Talk for the saved conversation. In Logs, compare transcription, time to first output and response latency. Human review scores are separate from measured recognition error and latency.
See troubleshooting if audio connects without a reply, or callers cannot interrupt.