Test and optimize voice behavior
Test behavior, audio, actions, and lifecycle separately. A pleasant sample conversation is not sufficient evidence for production readiness.
Browser test behavior
Select Test in the agent editor. Wirevox first attempts to flush pending draft changes, then starts a browser voice test using the saved draft.
A browser test:
- requires microphone permission;
- consumes workspace credits;
- does not require a published version or assigned phone number;
- can display a live transcript; and
- can execute connected tools for an owner test.
Because real actions are possible, use safe test inputs and verify the external system afterward.
Build a scenario matrix
At minimum, test:
| Scenario | Expected evidence |
|---|---|
| Normal request | Correct answer or completed primary outcome. |
| Ambiguous request | One focused clarification at a time. |
| Unknown fact | No invented answer; correct fallback or transfer. |
| Interruption | Agent stops or continues according to configured sensitivity. |
| Background and filler speech | No accidental interruption or false input. |
| Silence | Prompt and hangup occur at intended thresholds. |
| Function success | Inputs confirmed and returned result stated accurately. |
| Function failure or empty result | No false success; safe recovery path. |
| Knowledge hit and miss | Grounded answer on hit and explicit uncertainty on miss. |
| Sensitive or disallowed request | Boundary and escalation behavior followed. |
| Long conversation | Call remains coherent and respects the configured duration limit. |
Tune one layer at a time
- Fix factual or policy mistakes in the prompt, variables, knowledge, or function rules.
- Fix recognition mistakes with language, speech-recognition, and key-term settings.
- Fix pacing with response speed, speaking speed, and prompt concision.
- Fix turn-taking with interruption and filler-word controls.
- Fix tone with voice selection and delivery controls.
- Retest the same scenario after each change.
Changing several layers together makes regressions difficult to attribute.
Inspect evidence
Do not evaluate only by listening. Compare the transcript with what you heard, confirm function inputs and external results, inspect the call outcome and summary when present, and review credits charged.
Speech recognition can be wrong even when the model's reasoning is correct. A correct transcript paired with a wrong response points instead to instructions, knowledge, or model behavior.
Validate the live path
After draft testing passes:
- Publish a named version with a useful note.
- Confirm the intended number is assigned and the agent is Active.
- Place a new real call from an external phone.
- Repeat the highest-risk scenarios.
- Inspect the new call log and cost.
This second pass catches version, carrier, routing, audio, and permission differences that a browser test cannot prove.