Skip to main content

Test and troubleshoot speech

Test the complete audio path, not only the voice preview. A production result depends on the caller's microphone, phone network, speech recognition, turn detection, model response, text-to-speech provider, and playback connection.

Minimum test set​

Run the same cases after every material speech change:

  1. Opening greeting and a normal request.
  2. A caller with a different accent or speaking pace.
  3. Background noise and a weak connection.
  4. A long name, address, email, date, time, amount, and reference number.
  5. A pause in the middle of a sentence.
  6. An interruption while the agent is speaking.
  7. A function call with required details and confirmation.
  8. Every enabled language and one bilingual switch.
  9. A sensitive or frustrated caller.
  10. Silence, voicemail, and a natural close.

Use browser testing for rapid draft iteration, then place a real phone call before publishing. Review the transcript, the audio, action inputs and results, latency, and credit usage together.

Diagnose common problems​

The transcript contains the wrong word​

  • Confirm the selected language and recognition provider.
  • Add only the specific domain term to Recognition key terms.
  • Test another Modular recognition provider with the same recording.
  • Separate recognition from pronunciation: if the transcript is right but the audio is wrong, use the pronunciation dictionary.

The agent cuts the caller off​

  • Increase Modular recognition patience, or select a more patient Realtime turn-taking profile.
  • Test whether short pauses, fillers, or background speech trigger a turn ending.
  • Keep the agent's own responses short so callers have clearer opportunities to speak.

The response is too slow​

  • Reduce an unnecessarily patient turn profile.
  • Compare provider and architecture latency with the same prompt and call.
  • Check whether long model reasoning, knowledge retrieval, or function execution is the actual delay.
  • Avoid judging latency from text generation alone; measure from the caller's final word to audible response.

The wrong voice speaks after a language switch​

  • Confirm the agent uses Modular architecture.
  • Verify both the secondary language and secondary voice are selected.
  • Confirm the secondary voice supports that language.
  • Test whether the transcript identified the caller's language correctly.
  • Remember that Realtime does not execute configured dual-voice routing.

An internal language tag is spoken​

Wirevox normally removes bilingual response tags before synthesis. Reproduce the problem in a new test, capture the call ID and transcript, and report it as a pipeline issue. Do not add language tags to the authored prompt as a workaround.

Delivery settings have little or no effect​

  • Check the capability label for the active architecture and voice provider.
  • Remove conflicting tone or speed rules from the system prompt.
  • Use a stronger setting temporarily to confirm that the control is reaching the provider.
  • If the control is guided, evaluate several responses rather than expecting exact deterministic output.

A custom voice cannot be selected​

  • Confirm the clone belongs to the current workspace.
  • Wait for training to finish and refresh its status.
  • Preview the voice before attaching it.
  • If training failed, use clean single-speaker samples in a supported format and retry with confirmed permission.

Change safely​

  1. Record the current architecture, model, voice, recognition provider, languages, delivery settings, and estimated credit rate.
  2. Change one major variable at a time.
  3. Run the fixed test set and compare results.
  4. Save the draft only when it improves the target metric without breaking another case.
  5. Publish with a descriptive version name and note.
  6. Monitor live calls for recognition corrections, repeated questions, interruptions, action errors, transfers, latency, and cost.

Escalation evidence​

When reporting a speech problem, include the workspace and agent, draft or published state, call or test ID, architecture, model, voice, language, recognition provider, approximate timestamp, expected behavior, actual transcript, and whether the issue was audible during the call. Do not include credentials or unrelated caller data.