Showing 10 of 10 voices
Alex
Daniel
Emily
James
Jessica
Lily
Olivia
Robert
Sam
Sarah
Which voice for which kind of call
There is no correct answer here and nothing in the platform enforces one — any voice can be attached to any agent. What follows is a starting point based on how each voice is characterised, to save you playing all ten against every script.
Calls people did not ask for
Payment reminders, collections, claim updates, anything with bad news in it. The calm and gentle voices — Sarah and Emily — carry a slower, lower-pressure delivery, which matters more here than in any other category. A voice that sounds like it is selling something makes a difficult call worse.
Calls people are waiting for
Appointment reminders, order status, support callbacks. Clarity beats character: Jessica and Olivia are the clear and crisp voices, and on a reminder call being unambiguous about a date and a time is the whole job.
Calls that need energy
Outbound prospecting, event follow-up, reactivation. Alex is the lively voice and Lily the bright one. Energy holds attention in the first five seconds, which is where most cold calls are lost.
Calls that need authority
Legal intake, finance, high-value B2B. James is the deep voice and Robert the polished one. Both read as more formal, which suits a conversation where the caller is assessing whether you are serious.
Calls that need warmth
Donor stewardship, patient follow-up, membership. Daniel is the warm voice and Sam the soft one. In a check-in call, the goal is for the person to stay on the line long enough to say something true.
Still not sure?
Play the samples above back to back with your own script in mind, then test the shortlist on a real call. A voice that reads well in a five-second sample does not always hold up across three minutes of conversation, and the only way to find that out is to have the conversation.
Languages and accents
Being straight about this is more useful than a long list would be: the predefined library is English. All 10 voices above are English voices, and if you need your agent to speak another language, the route is a clone rather than a different entry in this list.
Every voice record carries a language field as a BCP-47 tag — the same tag scheme that distinguishes en-US from en-GB — so a cloned voice reports the language it was trained on, and your integration can filter on it client-side once the catalog contains more than one.
Agent configuration and voice are separate settings. An agent can be prompted in one language while speaking with a voice trained on another, which usually sounds wrong — match them deliberately rather than by accident.
How a voice gets attached to an agent
Voice is a property of the agent, not of the script. Templates in the agent catalog ship with no voice attached at all — you pick one when you clone a template into your workspace, which is why the same screening script can sound like a different company in two different accounts.
Listing voices over the API
GET /v1/voices returns every voice your API key can use: the predefined library above plus any of your own clones that have reached the ready state. It takes no query parameters — the scope of the response follows from the key you present, which needs the voices:read scope.
curl https://api.rexa.ai/v1/voices \
-H "Authorization: Bearer vk_live_..."The response is a data array plus an updated_at timestamp. Each entry carries:
| Field | Type | Notes |
|---|---|---|
| id | string | Pass this as `voice_id` on `POST /v1/calls` or in an agent’s configuration. |
| name | string | Display name for the voice. |
| language | string | BCP-47 / ISO 639-1 tag the voice was trained on. |
| gender | string | null | Speaker gender where known. Null for non-binary or unspecified clones. |
| tier | "standard" | "premium" | Pricing tier recorded against the voice. |
| sample_url | string | null | Audio preview, when one has been published for that voice. |
| is_cloned | boolean | False for stock voices; true once one of your consent-verified clones is ready. |
Take the id of the voice you want and pass it as voice_id when you dispatch a call or open a browser room. Authentication for both is covered in the authentication guide.
What the voice costs
The voice your agent speaks with is included in the per-minute rate rather than priced by voice. Every voice in the predefined library is on the standard tier, so switching between them changes how the agent sounds and nothing on the invoice. See pricing for what a minute covers.
Cloning a voice, and the consent that gates it
When none of the predefined voices is right — a founder’s voice, a named brand presenter, a language the library does not cover — you can clone one. A clone is not something you can create quietly: the flow is built so that a voice cannot reach a live call without a recorded consent statement from the person being cloned.
Upload samples and a scripted consent recording
The speaker records a fixed consent script alongside the voice samples. The clone stays in a pending state until both are present and long enough to work with.
Human review
The consent recording is reviewed before training starts. A clone that fails verification is rejected — the flow will not fall through to training on its own.
Training, then ready
Once the clone reaches the ready state it appears in GET /v1/voices alongside the predefined voices, flagged with is_cloned set to true, and can be attached to an agent like any other.
Revocable by the speaker
The person whose voice it is can withdraw consent at any time. Revocation deletes the clone and all derived assets within 24 hours.
Obtaining the speaker’s consent for the use you intend remains your responsibility as the customer — the platform enforces the verification step, not the legality of your use case. The full position is on the compliance page, and the security posture around stored audio is on the security page.
Hear one on a real call
A sample tells you the timbre; a conversation tells you whether it holds up. Pick an agent template, attach a voice, and talk to it in a browser tab — or read the FAQ first.