> ## Documentation Index
> Fetch the complete documentation index at: https://docs.ultravox.ai/llms.txt
> Use this file to discover all available pages before exploring further.

# Understanding VAD

> Learn how Ultravox's multi-model VAD system works and when to adjust voice activity detection parameters.

Voice Activity Detection (VAD) determines when a user is speaking and when they've finished their turn. Understanding how Ultravox's VAD system works will help you make informed decisions about when and how to adjust settings.

## How Ultravox VAD Works

Ultravox Realtime uses a multi-layered approach to VAD which results in low latency and an overall experience that feels fluid and natural.

### Traditional VAD Model

The foundation is a classic VAD model that analyzes audio frames (32ms each) to predict "speechiness" - whether each frame contains human speech. This model is intentionally aggressive, meaning it has a low threshold for detecting potential speech.

### Neural VAD

Our neural VAD model makes intelligent predictions about conversation state and turn-taking patterns. It understands context like:

* Whether the user is likely finished speaking based on conversation flow
* Typical pause patterns in natural dialogue
* The difference between a thoughtful pause and the end of a turn

### Noise Cancellation & Audio Processing

Additional models handle:

* Background noise filtering
* Echo cancellation
* Audio quality enhancement
* False positive reduction

## VAD Parameters Explained

Ultravox exposes several VAD parameters that control this behavior. These parameters are exposed via the [vadSettings](/api-reference/calls/calls-post#body-vad-settings) object that can be set when creating new calls or call stages.

The following settings are available:

### `turnEndpointDelay`

<ResponseField name="turnEndpointDelay" type="string" default="0.384s">
  The minimum time the agent waits before responding after the user appears to stop speaking.

  Only multiples of 32ms are meaningful (anything from 1ms - 31ms produces the same result).

  **Trade-offs:**

  * **Shorter delays** → Leads to faster responses, but more likely to interrupt users mid-thought.
  * **Longer delays** → Makes agent less eager to respond (i.e. interrupt the user), but perceived as slower or less responsive.

  ```js Adjusting turnEndpointDelay theme={null}
  vadSettings: {
    turnEndpointDelay: "0.384s"  // 12 VAD frames at 32ms each
  }
  ```
</ResponseField>

### `minimumTurnDuration`

<ResponseField name="minimumTurnDuration" type="string" default="0s">
  The minimum duration of user speech required to be considered a valid turn.

  **Trade-offs:**

  * **Shorter durations** → Captures very short user audio.
  * **Longer durations** → Ignores very short user audio segments that might be noise but could ignore meaningful, short user responses like "yes" or "no".

  ```js Adjusting minimumTurnDuration theme={null}
  vadSettings: {
    minimumTurnDuration: "0s"  // Consider all user audio
  }
  ```
</ResponseField>

### `minimumInterruptionDuration`

<ResponseField name="minimumInterruptionDuration" type="string" default="0.09s">
  The minimum duration of user speech required to interrupt the agent when it's speaking. Similar to minimumTurnDuration but provides a higher threshold for interrupting the agent. Ignored if the value is less than minimumTurnDuration.

  **Trade-offs:**

  * **Shorter durations** → More sensitive to interruptions, may trigger on background noise.
  * **Longer durations** → Less sensitive to noise, but may miss legitimate interruption attempts.

  ```js Adjusting minimumInterruptionDuration theme={null}
  vadSettings: {
    minimumInterruptionDuration: "0.09s"
  }
  ```
</ResponseField>

### `frameActivationThreshold`

<ResponseField name="frameActivationThreshold" type="number" default="0.1">
  The threshold for considering an individual audio frame as containing speech (0.1 to 1.0).

  **Trade-offs:**

  * **Lower thresholds** → More sensitive to quiet speech, but more false positives.
  * **Higher thresholds** → Less sensitive to noise, but may miss quiet or distant speakers.

  ```js Adjusting frameActivationThreshold theme={null}
  vadSettings: {
    frameActivationThreshold: 0.1  // Very sensitive to potential speech
  }
  ```
</ResponseField>

## Best Practices for VAD Tuning

<Warning>
  <b>Start with Defaults</b>

  The default VAD settings work well for most applications. Only adjust them if you have specific, tested issues that can't be resolved through environmental or hardware improvements.
</Warning>

### The Safest Parameter to Adjust

If you venture into adjusting VAD settings, `turnEndpointDelay` is the safest parameter to modify. As noted in the [API reference](/api-reference/schema/call-definition#schema-vad-settings-turn-endpoint-delay), "there's nothing special about this value" - it's simply a starting point that works well in most scenarios.

### Making Changes Safely

1. **Change one parameter at a time** → Isolate the effects of each adjustment.
2. **Test thoroughly** → Use real users and realistic environments.
3. **Monitor the trade-offs** → Every improvement in one area may cause issues in another.

### Environmental Solutions First

Before adjusting VAD parameters, consider:

* **Audio quality** → Better microphones reduce VAD complexity.
* **Network quality** → Poor connections can affect VAD performance.

Remember: VAD tuning is a series of trade-offs. The system is designed to be intelligent and adaptive. Trust the defaults unless you have a compelling, well-tested reason to change them.