Test and troubleshoot¶
Test the path a caller uses: microphone or telephone audio through LiveKit and
the configured providers. Deterministic tests and check are useful gates, not
substitutes for a call.
Start a browser call¶
Open http://localhost:3000, allow microphone access, and connect. In one short
call confirm:
- the worker accepts the room and the greeting plays first;
- prepared caller data appears where expected;
- participant speech is understood in the active language;
- participant and assistant transcripts appear;
- response audio starts within the bot's latency target;
- qualifying barge-in stops obsolete assistant audio;
- one authorized business tool reports an honest result;
end_callcloses cleanly.
Use synthetic or approved test accounts. Do not put customer data in browser logs or bug reports.
Measure where time goes¶
Record these separately:
- end of participant speech → committed STT text;
- committed STT text → first model text;
- first safe model text → first TTS audio;
- end of participant speech → first audible response.
The fourth is caller-perceived latency. Compare the same phrase, region, route, and network. Streaming should overlap model output and TTS instead of waiting for a full answer.
Cover bot behavior¶
Test registered and unknown callers, every important prompt branch and tool, each language and switch direction, both silence nudges, a true and false interruption, configured fallback, transfer, tool/API error results, Judge output, transcripts, and the terminal result.
A development browser may combine only ready STT, LLM, TTS, language, and voice choices returned by the backend. Production has no selector. Confirm that unavailable or language-incompatible choices are disabled in the page and rejected by the token API.
Run a browser-room call before handoff. Before telephone traffic, run a separate inbound SIP call through the target number, trunk, dispatch rule, and transfer destination. Browser success does not prove SIP.
Capture safe evidence¶
When a call fails, record:
- safe call and room IDs plus timestamp;
- VAANI version, environment, and profile identity;
- lifecycle phase and the slow or failing latency segment;
- normalized provider/tool error;
- whether participant audio, transcript, assistant text, and assistant audio existed.
Never attach credentials, raw provider payloads, or customer transcripts. Use the symptom guide in Troubleshooting, then finish the production checklist.