Where it fits
Realtime voice · Transcription · Speech synthesis
Live and speech-generation models have different capabilities. Preview model availability can change; select a model shown in your account.
Configure the connection
Choose Realtime → Gemini for native audio. For separate stages, use stt_provider or tts_provider set to gemini and select the corresponding model.
Assistant configuration fragment. Merge this into your assistant; select current model and voice IDs from your account before saving.
{
"pipeline_mode": "realtime",
"realtime_provider": "gemini"
}Keep provider secrets in workspace credentials. The public MCP provider matrix describes capabilities; it does not connect your provider account.
Test this boundary first
Check interruptions, tool completion and long turns in the target language. Compare the browser transcript with the audio you actually hear.
Save a known working configuration and move one test route first. Inspect the call log before increasing traffic. Follow the staged migration guide.
Understand the bill
APIFleet adds $0.03 per connected AI minute, billed per second. Provider and carrier usage are additional. With your own keys, provider charges go directly to the provider. No monthly subscription or free call credits; prepaid top-ups start at $10.