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

# Agent Speech

> Control how readily your agent pauses for callers and how quickly it responds after they stop talking.

Two behaviors control how your agent shares the conversation:

* **Interruptibility** governs what happens while the agent is speaking: how readily it yields when the caller talks over it. This is a setting you choose.
* **Resumption** governs what happens after the caller finishes: how quickly the agent takes its turn. This is adaptive: the agent measures each caller's pace and matches it automatically. There is nothing to configure.

## Agent pauses

When a caller makes a sound while the agent is speaking, the agent does not cut off at the first sound. It keeps talking for a brief moment while it works out what the caller is doing, then reacts:

| The caller says | What it is | What the agent does |
| :- | :- | :- |
| "Uh-huh." | A short acknowledgment | Keeps talking. No pause at all. |
| "Stop for a second." | A short interruption | Finishes its current word, stops, and responds. |
| "Wow, that is really, really interesting." | A long acknowledgment | Pauses while the caller speaks, then picks up exactly where it left off. |
| "Hey, I have a question about that." | A long interruption | Finishes its current word, stops, and responds. |

Two things hold in every case:

* The agent always stops at a word boundary. It never cuts off mid-word.
* Acknowledgments never derail the agent. Short ones are spoken through; long ones get a pause, and the agent picks up where it stopped.

What counts as a "short" utterance scales with the interruptibility setting: an easy-to-interrupt agent pauses for shorter caller utterances, while a difficult-to-interrupt agent holds through more.

## Interruptibility

How readily the agent yields the floor when the caller speaks over it. Set with the `interruptibility` field (integer, `0`-`4`, default `2`).

| Setting | API value | Caller experience |
| :- | :- | :- |
| Always interrupt | `4` | Any caller speech stops the agent. Nothing is treated as an acknowledgment; every utterance gets a stop and a response. |
| Easy | `3` | Yields quickly. Only the briefest acknowledgments are spoken through. |
| Balanced | `2` | Default. Short acknowledgments do not stop the agent; real interruptions do. |
| Difficult | `1` | Holds the turn longer and treats more caller speech as acknowledgment. Good for noisy lines and chatty listeners. |
| Block interruptions | `0` | The agent ignores caller speech entirely while talking and always finishes what it is saying. |

## Adaptive resumption

> [!NOTE]
>
> Adaptive resumption is rolling out and may not yet be active in your environment.

How quickly the agent responds once the caller stops talking is measured per caller, per call. The agent tracks the caller's own rhythm (how long their speech runs and how long their mid-thought pauses last) and adjusts how long it waits before taking its turn:

* A caller who speaks quickly, without hesitating, gets near-instant responses.
* A caller who speaks deliberately, pausing mid-sentence to think or to read out a number, gets an agent that waits out those pauses instead of jumping in, up to about two and a half seconds.
* A caller who changes pace mid-call is matched within about fifteen seconds.

The adaptation runs continuously in both directions: slowing down earns patience, and speeding back up restores snappy responses.

## Where to set interruptibility

* **Calls with an agent**: pass `interruptibility` in the [`/v1/calls`](https://docs.bland.ai/api-v1/post/calls-agent) request body alongside `agent_id`. It applies to that call and overrides the agent's own configuration.
* **Calls with a one-off prompt**: pass `interruptibility` in the [`/v1/calls`](https://docs.bland.ai/api-v1/post/calls) request body alongside `task`. It applies to the whole call.

## Choosing values

| Symptom | Try |
| :- | :- |
| The agent talks over callers who want to say something | Easy interruptibility, or Always interrupt |
| The agent stops for background noise or brief acknowledgments | Difficult interruptibility, or Block interruptions |

Response timing needs no tuning: an agent that seems to jump in on a hesitant caller will settle within a few turns as the measurement catches up to the caller's pace.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.