Voice

Your agent can talk. Voice turns any agent into a real-time conversational partner — you speak, the agent listens, thinks, and replies out loud, streaming as it goes.

Where voice works

In the workspace

Start a live voice call from the chat composer. Speak naturally and the agent answers in real time, looking things up as you talk. Longer work, such as anything in its sandbox, goes to a background task that shows up in the chat.

Over the phone

Connect a phone number from your own Twilio account so customers can call your agent directly. On Business and up, the agent can move a caller to a person who joins from the browser (see Hand a conversation to a person).

Choosing a voice

Manage voices in Console → Voices, then assign one to an agent. Three voice types are available:

Type	What it is	Best for
Preset	A built-in, ready-to-use voice	Getting started quickly
Designed	A voice generated from a text description (e.g. “warm, calm, mid-range female”)	Matching a brand persona without a recording
Cloned	A voice created from a short sample of a person’s voice, plus that person reading a consent line aloud	Reproducing a specific person’s voice

Clone a voice

Cloning needs the Pro plan or above. Pro holds up to 3 cloned or designed voices and Business up to 10. Enterprise has no limit. Owners, admins and members can clone a voice, and only owners and admins can delete one. With member governance on, a member’s new voice stays pending until an owner or admin approves it.

A clone needs two recordings from the same person on the same microphone. The first is a short sample of their voice. The second is that person reading one fixed sentence aloud. Google makes the voice, and it compares the two recordings before it creates it.

Go to Console → Voices and choose Clone voice at the top of the page. A Clone a voice panel opens.
Under Name, give the voice a name, such as “Maya (founder)”.
Under Sample, choose Record or Upload a file. To record, press the microphone, read the text on the screen aloud at your normal pace, and press it again to stop. The sample must be 10 to 30 seconds long, and a recording stops by itself at 30. An uploaded file can be WAV, MP3, M4A or WebM.
Under Say the consent line, press the microphone and read this sentence aloud, word for word:

I am the owner of this voice and I consent to Google using this voice to create a synthetic voice model.

Press the microphone again when you finish. The sentence is shown on the screen, and the same person records it, on the same microphone, once the sample is ready. It is in English for now. The recording must be 2 to 20 seconds long and stops by itself at 20.
Tick the box that confirms you have the speaker’s consent to clone the voice and that it will only be used for your organization’s agents.
At the bottom of the panel, choose Clone voice. Cloning takes about a minute, and the voice shows Generating the voice… in the library until it is ready.

Clone voice stays off until the name, the sample, the consent line and the ticked box are all in place.

What is kept

Both recordings go to Google to create the voice, and Auteryn does not store either one. Google processes the recordings under its paid Gemini API terms, which say it does not use prompts, files or responses to improve its products, and that it logs prompts and responses for a limited time only to detect abuse and meet legal disclosure requirements.

The cloned voice itself is stored with Google, and deleting the voice in the Console asks Google to delete it too. Only agents in your organization can use the voice. Auteryn keeps the voice’s name and a record of who ticked the consent box and when. Voices you cloned before the consent line was added keep working, and you do not need to record it for them.

If a clone does not work

You see	What to do
“This sample is under 10 seconds.” or “This sample is over 30 seconds.”	Record or choose a sample between 10 and 30 seconds.
“That was under 2 seconds. Read the whole line.”	Read the whole sentence in one go.
“That recording can’t be used.” or “This audio can’t be used.”	Record again in a quiet room, or choose a WAV, MP3, M4A or WebM file.
“Microphone access denied”	Allow microphone access in your browser settings, then try again.
“Recording couldn’t start”	Check that a microphone is connected and not in use by another app, then try again.
“The consent recording wasn’t accepted.” (on the voice’s card)	Google could not match the consent line to the sample. Delete the voice and clone it again, with the same person on the same microphone reading the sentence word for word right after the sample.
“The recordings couldn’t be used. Record both again in a quiet room, then try again.” (on the voice’s card)	Background noise, or too little speech. Audio that is itself AI-generated is refused too.
“Voice creation is busy right now.” (on the voice’s card)	Wait a few minutes, then clone the voice again.
“Voice creation failed on our side.”, “Voice creation took too long.”, “Voice creation did not finish.” or “Voice creation is unavailable right now.” (on the voice’s card)	Clone the voice again in a moment.
“Voice cloning is at capacity right now.”	Try again later, or contact support if it keeps happening.

A voice that failed stays in the library and counts toward your plan’s limit until an owner or admin deletes it.

Setting up voice on an agent

Go to Console → Voices and add a preset, designed or cloned voice.
Open your agent, go to its Voice section, turn on Live voice and choose the voice. Choose Save changes.
Open the agent in the workspace and start a voice call from the composer. For phone calls, connect a number from your Twilio account.

How it sounds natural

Streaming responses — the agent starts speaking as soon as it has something to say, rather than waiting for the full answer.
Barge-in — you can interrupt the agent mid-sentence and it will stop and listen.
Your language — the agent answers in the language the caller speaks.
Lookups mid-call — while you talk, the agent looks things up (the web, a page you name, its Knowledge and, with the right Connections, your orders and payments, see below) and says what it found. Work that needs its sandbox runs as a background task, and the agent tells you when it’s done. On a phone call, background tasks stay off unless you turn on Phone callers can run background tasks in the agent’s Voice settings.

Orders, payments and texts on a call

If the agent has a Shopify or Stripe Connection, it can use it during a call. On a call it only reads: it can’t refund, cancel or change an order.

During a call the agent can	It needs
Look up an order: payment and delivery status, items, total and tracking	A Shopify Connection
Find the Stripe charges for an order it looked up, and point out a charge made twice	A Stripe Connection (on a phone call, Shopify as well)
Text the caller something to keep, such as a case or tracking number	A phone call to a number with a Twilio Connection

On a phone call the agent can’t see who is calling, so it checks:

It shares an order only when the caller rings from the phone number on the order or gives the email address on it, and after four lookups that don’t match it stops looking up orders for the rest of the call.
It checks payments only for an order it looked up on this call, and only for the person who bought it. A caller who is only the delivery contact hears where the parcel is, not what was paid.
It never reads out an email address, phone number or street address from an order.

A text goes only to the caller and comes from the number they called. It starts with the agent’s name. It has to be plain text of up to 320 characters, with no links, web or email addresses, or other phone numbers. An agent sends at most three texts in a call, and one number receives at most five a day.

Each lookup and text appears as a step in the call’s conversation, such as “Looked up order #9140 in Shopify”. To connect the services, see Featured MCP and Twilio & WhatsApp.

Tips

Keep instructions conversational — voice agents read instructions the same way chat agents do, but spoken replies work best when they’re concise.
For customer-facing phone agents, define how to handle silence, escalation, and “I didn’t catch that” cases in the instructions.
Voice is credit-intensive (30 credits/min in the browser, 54 credits/min over a phone number). See Pricing & Credits.