> For the complete documentation index, see [llms.txt](https://learning.contextqa.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://learning.contextqa.com/ai-agent-testing/voice-ai-testing/dead-air-and-latency.md).

# Test Voice AI Dead Air and Response Latency

Diagnose long silences and slow voice responses using measured audio criteria, complete recordings, controlled cases, and targeted regression tests.

> **Documentation owner:** [Deep Barot](https://contextqa.com/academy/) · **Last updated:** November 11, 2026

## Quick answer

To test dead air and response latency, enable the relevant measured voice criteria, define the acceptable timing policy, and inspect the recording around each failure. Separate missed short replies, turn-detection behavior, backend delay, and audio-delivery problems. Rerun the same case after a targeted change and report any remaining slow intervals.

## Key takeaways

* Define the measurement and acceptance limit before interpreting a timing score.
* Use the actual call audio to investigate silence; a transcript alone is insufficient.
* A shorter wait can improve responsiveness without fixing the underlying recognition problem.

## Establish a timing policy

Open the project’s **Metrics** page and review **Measured Response Latency** and **Measured Dead Air**. Enable the criteria required by your policy. Keep them separate from the LLM-judged **Dead Air** and **VAD Responsiveness** criteria, which assess behavior contextually.

A five-second silence limit can be an illustrative test policy, but it is not a universal ContextQA recommendation. State what counts as silence, how the interval is measured, and whether a clearly announced backend operation has a different allowed duration. Use the current criterion definition rather than inventing a formula.

## Reproduce the symptom

Create cases with a short name, a one-word affirmative response, a full-sentence response, and a correction. Keep the requested task and environment stable. Record where the caller finishes, whether the system acknowledges the input, and when the next meaningful response begins.

| Observation                                         | Investigation                                                                 |
| --------------------------------------------------- | ----------------------------------------------------------------------------- |
| A short “yes” is followed by a long wait            | Compare recognition and end-of-turn behavior with a longer affirmative phrase |
| Silence occurs only during booking                  | Check backend operation duration and the agent’s progress messaging           |
| Generated speech is clear but received audio is not | Compare available speech-engine output with the actual phone recording        |
| Transcript appears complete but timing looks wrong  | Replay the full recording and check segment alignment                         |

These observations narrow a hypothesis. They do not prove a particular component failed without supporting evidence.

## Change one cause at a time

Reducing a listening timeout can shorten the wait after missed input, but it does not necessarily make the agent recognize that input. Prompting a caller to say a longer phrase may help that interaction while leaving ordinary short replies broken. Test both forms after the change.

If backend latency dominates, test the action separately and check whether the agent sets an accurate expectation while waiting. Avoid treating filler speech as task completion or masking a failed operation with a pleasant response.

## Report improvement precisely

Consider an illustrative policy of a maximum five-second gap. If one repeat produces a 4.9-second maximum and another produces 7.1 seconds, the result is improved but not consistently passing. Preserve both outcomes; do not report the agent as fixed based only on the better run.

Retain the case, configuration, audio interval, metric reasoning, and any dependency timing available. Use repeated cases to investigate variation without claiming statistical reliability from a tiny sample.

For broader failure classification, use [voice troubleshooting](https://learning.contextqa.com/ai-agent-testing/voice-ai-testing/troubleshooting.md). For interruptions and corrections, use [multi-turn context tests](https://learning.contextqa.com/ai-agent-testing/voice-ai-testing/interruptions-and-context.md).


---

# Agent Instructions
This documentation is published with GitBook. GitBook is the documentation platform designed so that both humans and AI agents can read, navigate, and reason over technical content effectively. Learn more at gitbook.com.

## Querying This Documentation
If you need additional information that is not directly available in this page, you can query the documentation by asking a question.

Perform an HTTP GET request on the following URL with the `ask` and `goal` query parameters:

```
GET https://learning.contextqa.com/ai-agent-testing/voice-ai-testing/dead-air-and-latency.md?ask=<question>&goal=<user_goal>
```

`ask` is the immediate question: it should be specific, self-contained, and written in natural language.
`goal` is what the user is ultimately trying to achieve, the reason they need the answer. Sharing it helps GitBook give you a better, more relevant answer. A goal is most helpful when it describes the outcome the user wants rather than restating the question. For example, with `ask=how do I create an API token`, a goal like `automate deployments from our CI pipeline` lets GitBook tailor the answer to that use case.

The response will contain a direct answer to the question and relevant excerpts and sources from the documentation.

Use this mechanism when the answer is not explicitly present in the current page, you need clarification or additional context, or you want to retrieve related documentation sections.
