← Latest news
Voice AI3 min read

How to test Vapi voice agent interruptions

A practical testing scenario: distinguishing short acknowledgments from requests to stop, and ensuring customer corrections are captured correctly.

How to test Vapi voice agent interruptions
Editorial illustration

Settings to evaluate

Vapi documentation outlines the stopSpeakingPlan for managing agent interruptions. Setting numWords to 0 enables voice activity detection, while higher values set a threshold for transcribed words. The acknowledgementPhrases parameter defines phrases that should not trigger an interruption, and backoffSeconds controls the audio output pause after an interruption occurs. These are the specific parameters to evaluate within your scenario.

EDITORIAL PERSPECTIVES

What this means in practice

Practical scenario: delivery clarification

Consider a scenario where an agent lists delivery terms, and a customer first says "uh-huh" and then asks to change the address. The team should pre-define the expected reaction for each utterance: continue the explanation after a short acknowledgment signal and allow the customer to finish their correction. The test must include both situations, as well as similar phrases with different meanings, such as "yes, but the address is different."

Separately, verify the result in your order management system. A short acknowledgment signal should not be treated as final consent to record changes. Before saving a new address, the script must repeat it and receive explicit confirmation. This is a business logic requirement that the team must implement and validate independently.

How to compare configurations

Prepare a consistent set of utterances for different settings: quiet background, background chatter, short acknowledgments, and explicit corrections. Change only one parameter at a time. For each run, record whether the agent stopped, whether it captured the full correction, and what was ultimately recorded in the order.

Select the configuration based on the combination of these observations. Fewer interruptions might mean the agent is missing necessary corrections; faster interruption alone does not prove the address was recorded correctly. Repeat the same test suite after changing the voice model or speech recognizer before the next release.

What to record in the protocol

  1. Expected reaction to short acknowledgments and correction phrases.
  2. Configuration, background conditions, and the actual timing of the interruption.
  3. Completeness of the captured correction and the final address record.
  4. Results of re-testing after any change to the voice agent component.

Questions and answers

Why is counting agent interruptions insufficient?

You must distinguish between unnecessary interruptions and relevant ones. Evaluate whether the customer finished their correction and whether it was recorded correctly.

Should I change all parameters at once?

For comparative testing, change one parameter at a time while keeping utterances and background conditions constant. This makes it easier to identify the cause of behavioral changes.

Sources

VOICEIP / INTECRACY GROUP

From an update to a working scenario

Voice agents, telephony and integration with your business systems — using your own infrastructure or international platforms.

Explore our services ↗

Continue reading

All news ↗
VoiceIP · AI assistant

Let’s talk.

Ask about VoiceIP services and agents. Answers come from this website.

Ready when you are

Speak after the greeting. Your microphone stays on during the call. Up to 3 minutes.

You can speak or type. We can switch between Ukrainian and English.

Contact our team ↗

Google Gemini processes your messages and voice to respond. Privacy policy