Let visitors speak to the agent out loud and hear it reply, using the same knowledge and guardrails as text.
The Voice Assistant lets a visitor speak to the agent instead of typing, and hear the answer back. It is the same agent as the chat, answering out loud inside the widget on your site — there is no phone number and no inbound call.
Open Capabilities, then Voice Assistant.
The Voice Assistant is available on Growth and above. On Free and Starter the panel shows an upgrade prompt. See Plans.
Settings
- Voice — the voice your agent speaks with. Play it to hear it, or choose Change voice to browse the full list, filtered by gender, language and style.
- Greeting — an optional line the agent says when a call starts. Use Hear it to listen to it exactly as written. Keep it to one sentence.
- Minutes — minutes used this month against your allowance.
The preview beside the settings shows the widget in the middle of a call.
What the visitor sees
The visitor starts a call from the button in the message box. The chat stays on screen: the agent's replies write themselves into the conversation as it speaks, and a colourful glow rises from the bottom of the widget while the visitor talks. The bar at the top shows the call's length and carries mute and end call. The visitor can also type during a call, which is easier for an email address or an order number, and the agent answers out loud.
Same brain, different mouth
The Voice Assistant runs the same pipeline as text: your knowledge base, your instructions, your topic boundaries, your capabilities. It transcribes what the visitor said, runs the agent, and speaks the reply.
Two things are deliberately different:
- Replies are shorter. The agent is told to answer in one to three short sentences and never to read a URL aloud. That instruction sits after your custom instructions, so it survives them.
- Cards still work. If the agent shows a lead form, a scheduler, products or a handoff mid-call, it says so out loud and the card appears in the conversation, exactly where it would in text. Quick links appear too but are never spoken about.
A call ends after a minute of silence (the visitor sees a warning first) and lasts at most 30 minutes. Ending a call never loses anything: the visitor is back in the same conversation, with every message and card where it was.
On WhatsApp and Telegram, a voice note is transcribed and answered in text.
How it is metered
The Voice Assistant is billed in minutes of active speaking time — the visitor's speech plus the agent's, and nothing else. Silence, thinking time, network delay and dashboard previews do not count.
It is metered separately from your conversation pool.
A conversation that is already running is never cut off mid-sentence: the allowance is checked when it starts, not while it is in progress. Growth and Scale can opt in to overage, with a budget you set, to keep voice working past the allowance.
Full counting rules and allowances are in Usage and limits.