# Agent Settings: Voice

![Voice settings panel in the Edit Agent modal, showing the voice enablement toggle, the Voice and Voice call sub-tabs, three voice source cards (Browser, OpenAI, Google), and a voice picker](../images/agent-settings/agent_settings_voice.webp)

## Overview

The Voice panel controls whether an agent can speak, which voice it speaks with, and how voice calls behave. Its header describes it simply: "Choose the voice your agent uses and how voice calls work." When voice is enabled, users can talk to the agent out loud and hear it reply, like a phone call.

The panel is split into two sub-tabs. The **Voice** sub-tab (documented on this page) selects the voice the agent uses to read out replies in chat, while the **Voice call** sub-tab configures how live voice calls behave. Voice calls run as real-time WebRTC sessions powered by LiveKit.

To reach this panel, open the **Edit Agent** modal for any agent, stay on the **Configurations** tab group, and select **Voice** in the left sidebar. The sidebar also lists Settings, LLM, Screen Share, Prompts, Skills, Safety, Privacy, and Disclaimers.

## Target Audience

**Administrator** | **Agent Builder**

## Settings Reference

#### Voice enablement toggle
The toggle at the top of the panel sits next to the text: "Let people talk to this agent out loud and hear it reply, like a phone call. When on, pick the voice and fine-tune how calls behave below." Turn it on to expose the voice and call options; the sub-tabs below let you pick the voice and adjust call behavior.

#### Voice / Voice call sub-tabs
Two pill-style sub-tabs switch between voice selection and call configuration. **Voice** covers "The voice your agent uses to read out replies in chat." **Voice call** covers call style, spoken language, provider, and call voice — see the Voice Call page for details.

#### Voice source
Three selectable cards choose where the agent's speech comes from:

- **Browser** — "Speak through the user's own device. No setup needed." Uses the built-in speech synthesis of the user's browser, so no voice picker appears for this source.
- **OpenAI** — "High-quality voices from OpenAI." Selecting this source reveals a searchable voice picker below.
- **Google** — "Natural-sounding voices from Google." Also reveals a searchable voice picker for Google voices.

The selected card is highlighted with a blue border (OpenAI in the screenshot).

#### Voice picker ("Select a voice")
When the OpenAI or Google source is selected, a **Voice** field appears with a "Select a voice" trigger. Clicking it opens a modal picker where you can search, browse, and preview voices before choosing one (see the Voice Selector page). Once a voice is saved, the trigger displays the currently selected voice, with an inline play button to preview it without opening the picker.

#### Voice instructions
![Voice instructions card showing the style prompt with Edit and Copy actions, a character counter, and one-click example presets](../images/agent-settings/agent_settings_voice_instructions.webp)

Below the voice picker sits an optional **style prompt** — a free-form instruction for *how* the voice should deliver replies, as distinct from which voice says them. "Speak slowly in a warm, encouraging tone, like a patient tutor" is the shape of it.

The card carries **Edit**, which opens a rich-text editor, a **Copy** action, an information tooltip, a counter that caps the prompt at 1,000 characters, and one-click example presets — *Warm and encouraging*, *Calm and measured*, *Energetic and upbeat* — for when you want a starting point rather than a blank field.

![Edit Voice Instructions modal with a rich-text editor for the style prompt](../images/agent-settings/agent_settings_voice_instructions_edit.webp)

What the instruction does depends on the source: OpenAI treats it as speech instructions, Google as a synthesis prompt. Both read it as guidance on delivery, not as content to say.

#### Save voice
The **Save voice** button persists your voice source, voice selection, and voice instructions to the agent's settings. It is the save action for this sub-tab only; the Voice call sub-tab has its own Reset and Save controls.

## How to Use

#### Step 1: Open the Voice panel
Open the **Edit Agent** modal, keep the **Configurations** tab selected, and click **Voice** in the sidebar.

#### Step 2: Enable voice
Turn on the toggle at the top so people can talk to the agent out loud and hear it reply.

#### Step 3: Pick a voice source
On the **Voice** sub-tab, click one of the three source cards: **Browser** (no further setup), **OpenAI**, or **Google**.

#### Step 4: Select a voice
If you chose OpenAI or Google, click **Select a voice** to open the voice picker. Search or browse the list, press a play button to hear a sample, and click a voice to select it.

#### Step 5: Describe how it should speak
Optionally open **Voice instructions** and write how the voice should deliver replies — pace, warmth, formality. Start from one of the example presets if it is easier to edit than to draft. The Browser source speaks through the listener's own device, so a style prompt has nothing to act on there; it applies to the OpenAI and Google sources.

#### Step 6: Save
Click **Save voice**. To configure how calls themselves behave (call style, spoken language, call provider and voice), switch to the **Voice call** sub-tab.

## Behavior Notes

- **A voice is only sent when you pick one.** Leaving the picker untouched preserves whatever is already saved on the server rather than overwriting it — the same "absent means no change" contract applies to the voice instructions.
- **Clearing an instruction is deliberate.** Emptying the field and saving clears the stored prompt; not touching it leaves the existing one in place.
- **The instruction guides delivery, not content.** It shapes how replies are spoken; what the agent says is still governed by its prompts, tools, and guardrails.
- **1,000 characters is the ceiling.** The counter is a hard cap, not a suggestion.
