Tune speech and turn taking

Start from the defaults

Open the agent's Speech settings and Realtime transcription sections. The controls available depend on the selected engine. Managed realtime engines handle parts of speech generation and turn taking themselves; not every pipeline control applies to every model.

Change one setting at a time, save, and compare the same recording conditions. Use a headset in a quiet room before treating echo or a noisy laptop microphone as an agent problem.

Common adjustments

Pronunciation maps a written word to its spoken form. Follow the editor's one-word-to-spoken-form-per-line format. Use it for clinic names, addresses and staff names, then listen to the result.

Clinic vocabulary helps recognition of names and terms. Keep the list focused on likely caller phrases rather than pasting a whole document.

Reminder text controls what the agent says after silence; leaving it blank uses the caller's language where supported. Test that a quiet caller gets a useful prompt without repeated interruptions.

Speech thresholds and timing affect how quickly a turn starts or finishes. An aggressive setting may feel faster but cut off a caller who pauses. Validate slow speakers, short answers, background noise and barge-in.

Diagnose a slow turn

Separate the delay before transcription, reasoning, external tool execution and playback. Calendar or CRM requests can dominate a turn. A short contextual filler can acknowledge that work is in progress, but does not mean the action succeeded. The agent must still wait for a successful tool receipt before confirming a booking, cancellation or message.