After I switched my assistant to a Custom LLM and set model.metadataSendMode to "off" via API, my live voice assistant works normally, but all Voice Test Suite runs have a silent tester (“caller”)—the tester never speaks. Chat tests pass.
From docs: Test Suites spin up an AI tester that talks to your assistant; you can override it via testerPlan.
What changed right before it broke:
* I issued:
curl -X PATCH "https://api.vapi.ai/assistant/
" \
-H "Authorization: Bearer
" -H "Content-Type: application/json" \
-d '{"model": {"provider":"custom-llm","model":"
","metadataSendMode":"off","url":"https://api.openai.com/v1"}}'
(This follows the Vapi Custom LLM guidance for fine-tuned OpenAI models.)
Current behavior
Voice Test Suites: tester side is completely silent (no greeting, no utterances).
Live calls: assistant speaks and works as expected.
Chat Test Suites: pass (no audio path).
There are community threads about “silent” behavior in voice scenarios, but mine is specifically the tester going silent.
What I’ve tried
1. Removed tester override to use default tester:
curl -X PATCH "https://api.vapi.ai/test-suite/
" \
-H "Authorization: Bearer
" -H "Content-Type: application/json" \
-d '{"testerPlan": null}'
(Per API docs, testerPlan is an override; null should revert to defaults.)
2. Explicit tester override with a standard OpenAI model and explicit TTS/STT:
curl -X PATCH "https://api.vapi.ai/test-suite/
" \
-H "Authorization: Bearer
" -H "Content-Type: application/json" \
-d '{
"testerPlan": {
"model": { "provider": "openai", "model": "gpt-4o-mini-2024-07-18" },
"voice": { "provider": "elevenlabs", "voiceId": "
" },
"transcriber": { "provider":"deepgram", "model":"nova-2", "language":"en" }
}
}'
Still silent. (Docs confirm Voice Testing uses a speaking tester; this config should be sufficient.)