> For the complete documentation index, see [llms.txt](https://docs.borndigital.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.borndigital.ai/realtime-agent/voice-and-listening.md).

# Voice and listening

Choose the speech-to-speech model and voice, the transcription, and how the agent listens: interruptions, turn detection, silence and no-input.

The settings on the right side of the editor decide how the agent sounds and listens. They take effect with the next **Go live**.

## Model & Voice

**Essential** tab:

| Setting               | Options                                                                                                                                                                                                                     | Default                                              |
| --------------------- | --------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ---------------------------------------------------- |
| **Realtime provider** | **OpenAI**, **Azure** (Azure OpenAI), **Gemini**                                                                                                                                                                            | OpenAI                                               |
| **Realtime model**    | OpenAI and Azure: `gpt-realtime-2.1`, `gpt-realtime-2.1-mini`, `gpt-realtime-mini`, `gpt-realtime-1.5`, `gpt-realtime-2`. Gemini: `gemini-3.1-flash-live-preview`, `gemini-2.5-flash-live-preview`. Or **Custom model ID**. | `gpt-realtime-2.1` / `gemini-3.1-flash-live-preview` |
| **Reasoning effort**  | **Minimal**, **Low**, **Medium**, **High** – for `gpt-realtime-2.1` and `gpt-realtime-2`. Higher effort means better answers but longer pauses.                                                                             | Medium                                               |
| **Voice provider**    | **Realtime model voice** – the model speaks with its own voice; **ElevenLabs** – a separate ElevenLabs voice                                                                                                                | Realtime model voice                                 |
| **Voice model**       | OpenAI and Azure: alloy, ash, ballad, cedar, coral, echo, marin, sage, shimmer, verse. Gemini: Zephyr, Puck, Charon, Kore, Fenrir, Leda, Orus, Aoede.                                                                       | marin / Zephyr                                       |

<figure><img src="https://4261467870-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F6iQTvxgRZRwPS1NgIGEb%2Fuploads%2Fgit-blob-62fe06e1863cd17f005f396356452de51ba89a79%2Frealtime-model-voice.png?alt=media" alt="The Model and Voice card on the Essential tab: Realtime provider OpenAI, Realtime model gpt-realtime-2.1, Reasoning effort Medium, Voice provider Realtime model voice and Voice model marin"><figcaption><p>Model &#x26; Voice with the defaults</p></figcaption></figure>

With **ElevenLabs** as the voice provider, choose the ElevenLabs **Voice model** and a **Voice persona** (or enter a **Custom voice id** – a voice is required). The **Advanced** tab offers the **ElevenLabs endpoint** (**EU** keeps the data in the EU), a **Language code** such as `en` or `cs`, and **Voice settings JSON** for fine-tuning.

{% hint style="info" %}
Gemini cannot be combined with an ElevenLabs voice. Use the Gemini model's own voice, or another realtime provider.
{% endhint %}

### ElevenLabs Realtime

ElevenLabs Realtime is a speech-to-speech provider that answers with native ElevenLabs voices, in two models: **Max** (default, better tool calls) and **Mini**. It works in chat bubbles, with digital humans and on phone numbers whose **Conversation routing** goes through the voice platform; it cannot be combined with a separate ElevenLabs voice. An EU account keeps the data in the EU.

Not in Digital Studio yet: Born Digital switches it on for an agent on request; the setting in Digital Studio is coming.

### Protected greeting and goodbye

An agent can protect its opening and its goodbye on phone calls from interruptions: the caller is not heard until the greeting has played (up to a set number of seconds), and the goodbye always plays to the end. This now also works for phone calls that run through the voice platform, including ElevenLabs Realtime. Calls tested in the browser are not affected.

Not in Digital Studio yet: Born Digital switches it on for an agent on request; the setting in Digital Studio is coming.

## Transcript

The conversation is transcribed separately from the model, for logs and analytics.

| Setting             | Options                                                                                                    | Default                  |
| ------------------- | ---------------------------------------------------------------------------------------------------------- | ------------------------ |
| **Model**           | `gpt-4o-mini-transcribe`, `gpt-4o-transcribe`, `gpt-4o-transcribe-diarize`, `whisper-1`                    | `gpt-4o-mini-transcribe` |
| **Language**        | **Auto detect** or one of 57 languages. Set it when all callers speak one language – it improves accuracy. | Auto detect              |
| **Prompt**          | Optional hints for the transcription, for example product names.                                           | empty                    |
| **Noise reduction** | **Near field** (phone held to the ear), **Far field** (speakerphone), **None**                             | Near field               |

<figure><img src="https://4261467870-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F6iQTvxgRZRwPS1NgIGEb%2Fuploads%2Fgit-blob-7c3ad98e69039d3123a27f195f4790fea8f2cd99%2Frealtime-transcript.png?alt=media" alt="The Transcript card with Model gpt-4o-mini-transcribe, Language Auto detect, an empty Prompt and Noise reduction Near field"><figcaption><p>Transcript settings</p></figcaption></figure>

## Listening

**Behaviour** tab:

| Setting                              | What it does                                                                                                                                                                          | Default |
| ------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | ------- |
| **Caller can interrupt**             | The caller can speak while the agent talks and the agent stops to listen.                                                                                                             | off     |
| **Agent silent on background noise** | The agent waits instead of reacting to noise or to sounds that are not meant for it.                                                                                                  | off     |
| **Idle timeout (ms)**                | If the caller says nothing for this long after the agent's answer, the agent continues the conversation. 5,000–30,000 ms (default value 10,000). Only with **Normal** turn detection. | off     |
| **No-input reminder (s)**            | If the caller is silent for this many seconds, the agent asks whether they are still there. 5–120 s (default value 15).                                                               | off     |
| **No-input hangup**                  | With the reminder on: the agent ends the call after **Max. consecutive no-inputs** reminders without an answer (1–10, default 3). Phone calls only.                                   | off     |

Use either the idle timeout or the no-input reminder – they cancel each other out, and only the reminder can end the call.

<figure><img src="https://4261467870-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F6iQTvxgRZRwPS1NgIGEb%2Fuploads%2Fgit-blob-f397bf5e9b30ade0269a13d539b59688b7ba4bb4%2Frealtime-listening.png?alt=media" alt="The Listening card on the Behaviour tab with the switches Caller can interrupt, Agent silent on background noise, Idle timeout (ms) with a slider from 5 000 to 30 000, and No-input reminder (s) with a slider from 5 to 120"><figcaption><p>Listening: the Behaviour tab</p></figcaption></figure>

**Turn detection** tab – how the agent recognises that the caller has finished speaking:

* **Normal** (default) – the caller has finished after a pause:
  * **Threshold** (0–1, default 0.5) – how loud speech must be to count; raise it in noisy environments,
  * **Prefix padding (ms)** (default 300) – audio kept from before the speech started,
  * **Silence (ms)** (default 500) – the pause that ends the caller's turn; raise it if the agent interrupts callers who are thinking.
* **Semantic** – the model judges from the meaning whether the caller has finished. **Eagerness** (**auto**, **low**, **medium**, **high**) sets how quickly it answers.
* **Disabled** – no automatic turn detection.

<figure><img src="https://4261467870-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2F6iQTvxgRZRwPS1NgIGEb%2Fuploads%2Fgit-blob-b526a8f0573b5136cb1d2c838adacdb20933385c%2Frealtime-turn-detection.png?alt=media" alt="The Turn detection tab with the types Normal, Semantic and Disabled, and the Threshold, Prefix padding (ms) and Silence (ms) sliders of the Normal type"><figcaption><p>Listening: the Turn detection tab</p></figcaption></figure>

## Limits

* A call can last at most **20 minutes**.
* The agent cannot transfer a call to an operator – see [MCP tools](https://docs.borndigital.ai/realtime-agent/mcp-tools.md#built-in-behaviour).
