Skip to main content

Test and troubleshoot call behavior

Test behavior as a sequence of audio events. A good transcript alone does not prove that the caller was heard at the right time, that playback stopped, or that the phone call ended correctly.

Minimum regression set​

  1. The normal opening greeting.
  2. The caller begins speaking during the greeting.
  3. A short backchannel such as “mhm” while the agent speaks.
  4. A deliberate interruption such as “No, Tuesday.”
  5. A mid-sentence thinking pause.
  6. Silence long enough to trigger the presence check.
  7. Continued inactivity through automatic hang-up.
  8. A call near the configured duration limit.
  9. Successful and failed function execution.
  10. Successful transfer and unavailable-transfer fallback.
  11. A normal resolved closing.
  12. Every enabled language and channel-specific flow.

Run fast iterations in the browser, then repeat timing, transfer, and hang-up cases over a real phone connection. Browser transport does not reproduce every telephony delay or audio artifact.

Diagnose by symptom​

The agent answers before the caller finishes​

  • Modular: increase Response Speed patience or compare recognition providers.
  • Realtime: select a more patient Voice Delivery turn-taking profile.
  • Confirm the issue is turn completion, not the inactivity timer.
  • Test the same pause with a fixed recording or script.

The caller cannot interrupt​

  • Modular: lower Words to Interrupt and verify the caller's words are not in Filler Words to Ignore.
  • Realtime: confirm provider-native interruption events are arriving; Modular thresholds do not apply.
  • Check for low input level, echo, or poor transcript confidence.
  • Keep agent responses short so successful interruption is not the only route into the conversation.

Backchannels repeatedly stop playback​

  • Modular: add only the exact harmless acknowledgement to Filler Words to Ignore.
  • Confirm that the phrase is not also used as a meaningful command in the workflow.
  • Realtime: evaluate the provider turn-taking profile; the Modular filler list is inactive.

The silence prompt repeats or happens while someone is speaking​

  • Capture the call ID and timestamps.
  • Check whether caller or agent activity was recognized.
  • Compare browser and phone behavior.
  • Confirm Silence Timeout is not being confused with Response Speed.

Wirevox intends one silence prompt per inactivity period; caller activity resets that state.

The call hangs up too early​

  • Compare the event time with Hangup Timeout and Limit Call Duration.
  • Remember that a disabled custom duration still has the 40-minute safety cap.
  • Check whether the transfer function intentionally disconnected the AI media path.
  • Review the telephony hang-up cause in the call record when available.

The transfer does not happen​

  • Confirm a Call Transfer function is attached to the agent draft.
  • Confirm the system prompt names that exact function and supplies its trigger.
  • Validate the destination phone number or voice-agent ID.
  • Use blind transfer for phone destinations; warm transfer requires a Wirevox voice agent.
  • Inspect the function result and call logs before changing prompt wording.

The wrong greeting plays​

  • Confirm whether the call is inbound, outbound, or an agent-to-agent handoff.
  • Check the inbound Custom Greeting separately from the system prompt.
  • Remember that a handoff uses transfer context instead of replaying the ordinary greeting.
  • Verify whether the live call uses the published version while browser testing uses the draft.

Voicemail or no-answer fallback does nothing​

Do not tune the prompt. General voicemail detection/drop and transfer no-answer fallback are not currently executed production controls. Use an attached message-taking or follow-up function and test that path directly.

Evidence for escalation​

Include the workspace, agent, draft or published state, call or test ID, timestamp, architecture, model, recognition provider, voice, relevant control values, transcript, expected audio event, actual audio event, function result, and telephony hang-up cause when available. Remove unrelated caller data and never include credentials.

Publish safely​

  1. Record the current configuration and a baseline result.
  2. Change one behavior control at a time.
  3. Run the fixed regression set.
  4. Review latency, successful interruptions, false interruptions, transfers, outcomes, duration, and credits.
  5. Publish with a descriptive version name and note.
  6. Monitor new calls before making another large timing change.