> ## Documentation Index
> Fetch the complete documentation index at: https://docs.voiceinfra.ai/docs/llms.txt
> Use this file to discover all available pages before exploring further.
# Call Behaviour: Interruptions, Response Timing and Call Limits
> Choose how your agent handles interruptions and how long it waits before replying — with presets for patient, balanced, and responsive conversation, plus call duration, silence, voicemail and privacy settings.
**Call Behavior** controls how your agent handles interruptions and how it responds when the caller stops speaking. These settings decide whether an agent feels snappy or unhurried, whether it gets talked over, and whether it mistakenly stops for a cough.
You will find it on the agent's **More Settings** tab, under **Advanced settings**.
These settings apply **globally across all workflow nodes**. Individual nodes do not have their own interruption or endpointing settings.
## Start with a preset
Most agents never need to touch the individual fields. Pick the preset that matches your callers and you are done.
| Preset | Best for |
| - | - |
| **Patient** | Medical, legal, and elderly callers. Never rushes. Waits for long pauses. |
| **Balanced** | General customer support. **The default for new agents.** |
| **Responsive** | Sales and quick back-and-forth. Fast replies, tolerates natural interruptions. |
| **Custom** | You set individual fields yourself. |
### What each preset sets
| Setting | Patient | Balanced | Responsive |
| - | - | - | - |
| Minimum speech duration | 0.8s | 0.4s | 0.2s |
| Minimum words to interrupt | 2 | 1 | 0 |
| False interruption timeout | 2.0s | 2.0s | 1.0s |
| Minimum response delay | 0.6s | 0.3s | 0.15s |
| Maximum wait time | 4.0s | 2.5s | 1.5s |
| Pre-generate audio | No | No | Yes |
| Skip for utterances longer than | 5.0s | 10.0s | 10.0s |
All three presets also share these values: detection mode **VAD**, **Resume after false interruption** on, **Discard audio during locked speech** on, endpointing **Dynamic**, and **preemptive responses** on.
Your agent shows **Custom** as soon as any field differs from a preset — including a change you made yourself. That is expected. It means the agent no longer matches a preset exactly.
## Interruption
What it takes for the caller to stop the agent mid-sentence.
| Setting | What it does | Range |
| - | - | - |
| **Detection mode** | How an interruption is recognised. | VAD (Voice Activity) |
| **Minimum speech duration** | How long the caller must speak before the agent stops. Higher means harder to interrupt. | 0 – 5s |
| **Minimum words to interrupt** | How many words the caller must say before the agent stops. | 0 – 20 words |
| **Resume after false interruption** | If the agent stopped for a cough or a short sound, whether it picks up where it left off instead of restarting. | On / off |
| **Discard audio during locked speech** | During non-interruptible speech, such as a goodbye message, throw away anything the caller says. Prevents buffered audio from triggering a response once the lock releases. | On / off |
| **False interruption timeout** | How long to wait before deciding an interruption was a false alarm. **Set to 0 to disable recovery entirely.** | 0 – 30s |
**Minimum words to interrupt** does not apply to realtime agents — there is no separate speech-to-text pass for the model to count words from.
## Endpointing
Endpointing decides how long the agent waits after the caller stops speaking before it replies. This is the single biggest lever on whether an agent feels natural.
| Setting | What it does | Range |
| - | - | - |
| **Detection mode** | **Dynamic** adapts to the rhythm of the conversation. **Fixed** always waits exactly the minimum delay. | Dynamic / Fixed |
| **Minimum response delay** | The fastest the agent will reply after the caller stops. Lower is snappier but may cut callers off mid-thought. | 0.05 – 5s |
| **Maximum wait time** | The longest the agent will wait through a pause before replying. Higher is more patient with slow speakers. | 0.1 – 30s |
The minimum delay must be less than the maximum wait time.
If callers complain about being interrupted, raise the **minimum response delay** first. If the agent feels sluggish, lower it before touching anything else.
## Preemptive generation
Lets the agent start preparing a reply before it is certain the caller has finished, which shortens the perceived pause. If the caller keeps talking, the speculative reply is thrown away.
| Setting | What it does |
| - | - |
| **Enable preemptive responses** | Starts LLM inference before the caller has fully finished speaking. |
| **Also pre-generate audio** | Runs text-to-speech speculatively alongside the model for the biggest latency saving. Costs slightly more, because a wrong guess wastes the speech generation. |
| **Skip for utterances longer than** | If the caller has been speaking longer than this, do not speculate — long utterances are unlikely to be finished. Lower values save cost; higher values speculate more. Range 1 – 60s. |
## OpenAI Realtime Turn Detection
Agents running an OpenAI realtime model use a different mechanism: the model detects speech boundaries itself. The interruption settings above still apply for stopping the agent's speech, but endpointing and preemptive generation do not.
| Setting | What it does |
| - | - |
| **Detection type** | **Semantic VAD** uses a classifier that judges completion from the words spoken. **Server VAD** uses silence duration. **None** disables turn detection entirely. |
| **Eagerness** | How quickly the model responds. **Low** lets callers take their time, **High** responds as soon as possible, **Auto** is equivalent to Medium. |
| **Activation threshold** | How loud audio must be to count as speech. Higher is better in noisy environments. |
| **Prefix padding** | How much audio before detected speech is kept, preserving context at the start of a sentence. 0 – 5000 ms. |
| **Silence duration** | How much silence marks the end of a turn. Shorter detects turns faster but may cut speakers off. 100 – 10000 ms. |
| **Create response after speech detection** | Whether the model replies automatically once the caller stops. Turn off to control replies yourself. |
| **Allow interruption of agent responses** | Whether the caller can cut the agent off mid-response. |
Setting detection type to **None** disables automatic speech boundaries. OpenAI will not detect when the caller starts or stops speaking, and you become responsible for turn detection.
## Call duration and ending
| Setting | What it does | Range |
| - | - | - |
| **Max Call Duration** | Hard limit on a call. The agent warns the caller before reaching it, then ends the call. | 1–15 minutes (default 3) |
| **Max Call Duration Warning Message** | What the agent says as the call approaches the limit. | — |
| **End Call Function** | Lets the agent hang up on its own once the conversation is finished, using the phrases and message below. | On / off |
| **End Call Phrases** | Phrases that should end the call, such as *goodbye* or *that's all*. | — |
| **End Call Message** | The final line the agent says before hanging up. | — |
The limit counts connected talk time rather than ringing, so a slow answer does not eat into the caller's allowance.
**End Call Phrases are guidance for the model, not exact matching.** They are passed to the agent as a description of when hanging up is appropriate, so it recognises variations naturally rather than needing an exact string.
## Silence and timeouts
Separate from endpointing: if the caller goes quiet for a long time, the agent checks whether they are still there.
| Setting | What it does | Range |
| - | - | - |
| **Silence Timeout Duration** | How long to wait in silence before checking on the caller. | 5–300 seconds (default 30) |
| **Timeout Message** | What the agent says when the silence timeout fires. | Up to 500 characters |
| **Maximum Timeout Attempts** | How many times the agent will check before ending the call. | Default 3 |
Each time the caller speaks, the timer and the attempt count reset. If the agent reaches the maximum number of attempts without a reply, it ends the call.
## Voicemail handling
For outbound calls, the agent can tell the difference between a person answering and an answering machine.
| Setting | What it does |
| - | - |
| **Voicemail Message** | What the agent says if it reaches voicemail. Leave blank to hang up silently instead. |
| **Voicemail Detection** | Whether the agent checks for voicemail at the start of outbound calls. |
Voicemail detection applies to **outbound** calls only. An inbound caller is a person by definition. Detection runs before the agent greets, so a voicemail does not trigger the opening message.
If the agent reaches a mailbox that cannot accept a message, it ends the call without speaking. Both outcomes are recorded on the call log so you can measure live-answer rates.
## Recording and Privacy Mode
Both settings live under **Privacy** on the **More Settings** tab.
| Setting | What it does |
| - | - |
| **Privacy Mode** | Enables additional privacy protections for handling sensitive information. **When enabled, audio recording is automatically disabled.** |
| **Audio Recording** | Records the call audio and makes it available from the call log. Disabled while Privacy Mode is on. |
If you enable recording, make sure you comply with the consent and notification rules that apply where you and your callers are. Recording conversations without the required notice can be unlawful.
Recordings are stored securely and played back from the call log. Recordings stop when a call is transferred.
## Background audio
Quiet ambience played underneath the conversation, so the agent does not sound like it is speaking from an empty room.
| Setting | What it does |
| - | - |
| **Background Audio** | The ambience to play. |
| **Background Audio Volume** | How loud it sits behind the conversation. |
Background audio starts when the call connects and stops when it ends.
## First message mode
The opening message can be delivered in two ways:
* **Static** — the agent says your first message exactly as written.
* **Dynamic** — your first message is treated as an instruction and the model phrases it naturally for the moment.
Either way your text is the source of truth; dynamic mode changes the wording, not the meaning.
**Allow interruptions** controls whether the caller can cut into the **opening message** specifically. When off, the greeting always plays in full before the agent starts listening.
## Timezone
The agent's **Timezone** decides the date and time used by the built-in date and time variables in your prompts and first message. It does not affect business hours — agent behaviour is the same at every hour unless you build a workflow or transfer rule that checks the time.
Call logs and timestamps are always recorded in UTC.
## Caller's number passthrough
Controls whether the original caller's number is passed through when the agent transfers the call onward. Turn it on when the receiving party needs to see who is really calling; turn it off when the transfer should present the agent's own number.
## Frequently asked questions
**Balanced** unless you have a reason not to — it is the default for new agents. Move to **Patient** if callers are elderly, anxious, or explaining something complicated. Move to **Responsive** for sales and fast qualification calls.
Custom means the settings no longer match a preset exactly. Changing any single field — or applying a preset and then editing one value — switches the label. Nothing is broken.
Raise **Minimum speech duration** and **Minimum words to interrupt** so a brief noise does not count as an interruption, and make sure **Resume after false interruption** is on so the agent carries on from where it stopped.
Raise the **Minimum response delay** and the **Minimum words to interrupt**. Both make the agent wait for more certainty before it starts speaking.
Lower **Minimum response delay** and **Maximum wait time**, and turn on **preemptive responses**. For realtime agents, adjust **Eagerness** instead — endpointing settings do not apply there.
No. Interruption, endpointing, preemptive generation, silence timeout, voicemail, recording, and background audio are all voice-only and have no meaning in a text conversation.
Endpointing is the short pause the agent leaves after someone stops speaking — fractions of a second to a few seconds. The silence timeout is the long wait that triggers a check-in — tens of seconds — and eventually ends the call.
It disables audio recording and switches the agent into its privacy-protective handling for sensitive information. The recording switch stays disabled until you turn Privacy Mode back off.
This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.