---
title: Voice Setup
summary: Connect LiveKit and a speech provider to turn on real-time voice for your agents.
date: 2026-09-09
---

# Voice Setup

Connect LiveKit and a speech provider to turn on real-time voice for your agents.

Voice needs two separate things connected before an agent can hold a spoken conversation: a transport (LiveKit) and at least one speech provider (Deepgram or ElevenLabs). Both live under **Settings → Voice**.

## 1. Connect LiveKit (transport)

LiveKit carries the actual audio between your browser and the agent. Enter your LiveKit project URL, API key, and API secret, then save. This is an install-wide setting: one LiveKit project serves every organization on this install, not one per organization.

## 2. Connect a speech provider (per organization)

Transcription and spoken replies are configured separately, and can use different vendors:

- **Deepgram**: transcribes English, Spanish, French, and Mandarin; speaks English, Spanish, and French.
- **ElevenLabs**: transcribes and speaks all four.

Add the provider's API key for each role you want to enable. You don't need both roles: voice works with transcription alone. See [Voice](/en/docs/features/voice) for what that looks like from the chat.

## 3. Test before you rely on it

Use **Test connection** before saving a new or changed credential. It probes the exact configuration you're about to save, not whatever is already stored, so a bad key doesn't get committed on top of a live install and interrupt anyone already in a voice session.

## 4. Vendor training opt-out

Both supported vendors let you opt out of having them train on your organization's audio, on by default. Some vendors only honor this on paid vendor plans: run **Test connection** after enabling it to confirm the vendor actually accepted the setting.

## Starting a session

Once transport and at least a transcription provider are connected, a microphone control appears in the agent conversation view. If it doesn't, check **Settings → Voice** for which piece is still missing. See [Voice Not Available](/en/docs/troubleshooting/voice-not-available) for what each specific gap means.

## Related

- [Voice](/en/docs/features/voice): what a voice conversation looks and behaves like
- [Voice Not Available](/en/docs/troubleshooting/voice-not-available): fixing a missing or disabled microphone control
- [Security](/en/docs/guides/security): how provider credentials are stored and scoped