---
title: Voice activity detection
description: "The Voice activity card under Advanced: a Mode preset for how long the agent waits before it answers, and the five thresholds behind it."
---

<Badge variant="accent">Voice agents</Badge>

Voice activity detection (VAD) listens to the caller's audio and marks where
speech starts and stops. Every downstream step, transcription, <Term>turn detection</Term> and
the silence timers, works from those marks. The **Voice activity** card under
**Advanced** sets them with a **Mode**, *how long the agent waits before it
answers*, and exposes the five thresholds behind it under **Advanced**.

:::note[The defaults suit most agents]
The card's own description says so, and it is right. Change one value at a time,
test with **Call**, and keep the version that improved the transcript.
:::

![The Voice activity card.](/media/build/voice-activity-card.webp)

## Configuring it

1. Open the agent and choose **Advanced** under **Build**.
2. In **Voice activity**, pick a **Mode**. To tune one threshold instead, open
   **Advanced** in the card and drag its slider or type into its field.
3. **Save**.

## Modes

A mode is a preset for all five thresholds. Nothing stores the mode itself: the
card reads it back from the values, so changing any threshold by hand shows the
mode as **Custom** (*The thresholds below no longer match a preset.*). An agent
that was never tuned reads as **Balanced**.

| Mode | For | Min speech | Min silence | Activation | Prefix padding | End-of-speech |
|---|---|---|---|---|---|---|
| **Fast** | Short, brisk exchanges: replies as soon as the caller pauses | 100 ms | 100 ms | 0.4 | 100 ms | 0 ms |
| **Balanced** | The recommended trade-off between responsiveness and interrupting | 100 ms | 200 ms | 0.5 | 200 ms | 100 ms |
| **Patient** | Callers who think mid-sentence: waits longer before replying | 200 ms | 600 ms | 0.6 | 300 ms | 500 ms |

## Reference

| Field | Default | Unit | Range | What it does |
|---|---|---|---|---|
| **Min speech duration** | 100 | ms | 100 to 3000, step 100 | Shortest sound considered speech. Lower = more eager |
| **Min Silence Duration** | 200 | ms | 100 to 3000, step 100 | Silence required to treat the user as done speaking |
| **Activation threshold** | 0.5 | | 0.1 to 0.9, step 0.1 | How loud or speech-like audio must be to count as speech. Lower = more sensitive, more false positives |
| **Prefix Padding Duration** | 200 | ms | 100 to 3000, step 100 | Audio prepended to detected speech, so the first syllable is not cut |
| **End-of-speech timeout** | 100 | ms | 0 to 3000, step 100 | Silence required before treating the turn as ended |

## What each one changes

- **Min speech duration** filters clicks and coughs. Raise it if the agent reacts to
  noise; lower it if it misses "yes" and "no".
- **Min Silence Duration** is the pause that ends a caller's speech. Raise it for
  callers who pause between phrases; lower it for faster replies.
- **Activation threshold** is the sensitivity. Raise it on noisy lines, or turn on
  [noise reduction](/build/voice-and-audio/noise-suppression) first and leave it
  alone.
- **Prefix Padding Duration** keeps the start of words. Rarely needs changing;
  raise it if transcripts lose first syllables.
- **End-of-speech timeout** is added after the silence before the turn is handed
  on. Together with **Min Silence Duration** it sets the floor of the agent's
  response time.

## Behaviour and limits

- Response latency is at least **Min Silence Duration** plus **End-of-speech
  timeout**, plus the model's own time. With the defaults these two add 300 ms.
- [Turn detection](/build/speech-pipeline/turn-detection) can delay a turn beyond
  what VAD decided but never shorten it.
- The **Max silence** timer under **Reminders** counts time in which VAD heard no
  speech; a threshold set too high can make a talking caller look silent.
- The settings apply to the browser **Call** test too.

## Verify

Start a call under **Testing**, **Call**, say short words and long sentences, and
watch the transcript: every utterance should appear once, complete, and the agent
should answer after a natural pause.

## Troubleshooting

- **The agent answers before I finish.** Switch the **Mode** to **Patient**, or
  raise **Min Silence Duration** in 200 ms steps; then turn on turn detection.
- **The agent is slow to answer.** Switch the **Mode** to **Fast**, or lower
  **End-of-speech timeout** first, then **Min Silence Duration**.
- **Background noise triggers the agent.** Enable denoising, then raise
  **Activation threshold** by 0.1.

## Related

- [Turn detection](/build/speech-pipeline/turn-detection)
- [Noise suppression](/build/voice-and-audio/noise-suppression)
- [Silence, reminders and call duration](/build/speech-pipeline/silence-reminders-and-call-duration)
- [Call quality metrics](/operate/call-quality-metrics)