Voice agents

A voice agent is the phone half of an AI agent. The instructions, knowledge and tools you set up in AI Agent are the same ones the caller hears, so an agent that answers in chat can answer on the line without being rebuilt.

This page covers the settings that only apply to speech. Everything else lives under AI Agent.

The Voice tab of an AI agent: the opening line, the speech-to-text and text-to-speech providers, and the turn-taking settings.
The Voice tab of an AI agent: the opening line, the speech-to-text and text-to-speech providers, and the turn-taking settings.

Putting an agent on a call

Add a Voice agent node to a call flow and pick the agent.

Providers

Three stages turn a call into a conversation: speech becomes text, a model decides what to say, and text becomes speech.

StageOptions
Speech to textaws
Text to speechaws, elevenlabs
Modelopenai, default gpt-4o

Opening the conversation

first_message_mode decides who speaks first.

ModeWhat happens
assistant_firstThe agent reads your greeting, word for word
assistant_first_generatedThe agent opens in its own words
user_firstThe agent waits for the caller to speak

Turn taking

turn_stop_mode decides how the agent knows the caller has finished speaking.

ModeHow it decides
smart_turnA classifier judges whether the thought is complete
speech_timeoutA fixed pause counts as the end of the turn

Interruptions and muting

Mute windows stop the agent hearing itself or the caller at particular moments: always, first_speech, until_first_bot_complete, function_calls.

first_speech and until_first_bot_complete describe the same window two different ways. The runtime combines them, so switching both on mutes more than either one asks for. Pick one.

Background sound

An empty line sounds dead to a caller, so the agent can play a room behind itself: office, call_center, keyboard_typing, hold_music.

Tools

Choose which of the agent’s tools it may use on a call. Two are built in for voice:

  • End call hangs up.
  • Transfer call passes the caller to a person.

Tools take time to run, so each one can say something while it works. The line is either a fixed sentence you write, or one the model improvises in context.

DTMF

Handoff