> 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/interruptions-and-context.md).

# Test Voice AI Interruptions and Multi-Turn Context

Evaluate barge-in, crosstalk, combined answers, and corrected details without confusing audio turn taking with business-context accuracy.

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

## Quick answer

Voice agents must handle interruptions while retaining the user’s final intended request. Test audio behavior with interruption and turn-taking criteria, and test meaning with Multi-Turn Context and task-specific rubrics. Review both the recording and the final action: an agent can yield politely yet still use an outdated date or name.

## Key takeaways

* Separate yielding and overlap behavior from retention of corrected information.
* Test combined details, repeated questions, and interruptions at meaningful points.
* Use matched text and voice cases to narrow failures without assuming their cause.

## Define the expected interaction

Write down when the agent should yield, which information must survive an interruption, and how it should resume. For example: during a booking readback, the caller interrupts to change the date. The agent must stop the old readback, acknowledge the correction, reconfirm the final details, and avoid creating a booking with the superseded date.

Enable relevant criteria from [the voice metric reference](https://learning.contextqa.com/ai-agent-testing/voice-ai-testing/metrics-reference.md), such as Agent Interruptions, Yields When Interrupted, and Turn Taking. Add Multi-Turn Context and a custom criterion for the final business values. Do not ask one broad “conversation quality” score to represent every requirement.

## Create complementary cases

| Case                    | Caller behavior                         | Assertion                                          |
| ----------------------- | --------------------------------------- | -------------------------------------------------- |
| Calm baseline           | Gives one detail at a time              | Agent completes the normal workflow                |
| Combined answer         | Gives name and service in one sentence  | Both values are retained                           |
| Mid-readback correction | Changes the date while the agent speaks | Old date is discarded and new date confirmed       |
| Clarification           | Asks the agent to repeat or explain     | Context survives the extra exchange                |
| Short interjection      | Says “no” or “wait”                     | Agent does not treat it as approval                |
| Repeated correction     | Corrects a value twice                  | Final value wins and earlier values are not reused |

Use a dynamic case to explore adaptive behavior. Use a static case to reproduce the particular phrase sequence that exposed a defect. Review the simulated caller’s behavior so an invalid or contradictory case does not become a misleading failure.

## Inspect both audio and meaning

Replay the interruption interval. Check whether speech overlaps, whether the caller was audible, and whether the agent yielded. Then inspect the later confirmation and backend action. The absence of crosstalk is not proof of correct context retention.

Compare the same request through text when the agent supports both modalities. If text succeeds and voice fails, investigate the speech path and its handoff to the agent. Do not immediately attribute the failure to recognition: timing, truncation, session handling, or the evaluation’s audio alignment can also affect the evidence.

## Retest without contaminating the fixture

Use an independent booking slot or reset fixture for each normal-success case. If a calm caller books the only available slot before an impatient caller runs, the second failure may be expected unavailability. That is different from losing multi-turn context. Preserve intentional collision scenarios as separate tests with their own acceptance criteria.

Finish by comparing requested, confirmed, and persisted values using [booking outcome checks](https://learning.contextqa.com/ai-agent-testing/voice-ai-testing/booking-outcomes.md), then apply [regression review](https://learning.contextqa.com/ai-agent-testing/agent-evaluation/runs-and-regression.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/interruptions-and-context.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.
