Nobody hires a receptionist sight unseen
You would not put someone on your main line without hearing them talk. Yet most AI phone products ask you to enter a card, wait for provisioning, and find out afterward what your business sounds like to a stranger.
We built it the other way around. You hear the voice, you talk to the agent, and then you decide.
Ten OpenAI voices, with audio you can actually play
The OpenAI realtime engine ships ten voices: marin, coral, sage, shimmer, and alloy on the female side, cedar, echo, ash, ballad, and verse on the male side. Every one of them has a real preview file you can play during signup. Not a description of the voice. The voice.
That roster is smaller than the list OpenAI publishes for text to speech, and that is deliberate. We probed each candidate against the live realtime model on 2026-07-18 and cut the ones the realtime API silently rejects. If a voice is in the picker, it works on a real call. Anyone who picked one of the dropped voices before gets migrated automatically.
Five Grok voices
The xAI engine adds five: rex, eve, ara, sal, and leo. These do not have preview files, and we are not going to fake one by running a text to speech render and calling it a preview. You hear a Grok voice by selecting it and placing a real call.
The whitelist that defines those five doubles as an injection guard, so a voice name arriving from anywhere other than the picker gets rejected rather than passed through to the provider.
When the choice gets made for you
Medical and dental practices, and any organization with a signed BAA on file, are locked to OpenAI. xAI does not offer a HIPAA BAA covering call audio, so the option is removed rather than left available with a warning.
Two more mechanics worth knowing. Browser test calls always run on OpenAI, so a Grok org testing in Agent Studio is hearing the OpenAI rendition of its agent, and the real phone line uses Grok as selected. And if the xAI key is unset in the environment, the bridge falls back to OpenAI rather than failing the call.
Agent Studio: talk to it, change it, talk again
Agent Studio is one screen. Your live phone number is in the corner, because this is not a sandbox, it is the same production configuration a real caller reaches.
You press the orb and talk to the agent in your browser over WebRTC. Or you type, if you are somewhere you cannot talk. Then you change something and go again:
- Playbook sets the greeting and the persona: warm, professional, or brief
- Knowledge takes a document, a menu, or a URL and the agent plays back what it just learned
- Voice switches the engine and the specific voice
The one guardrail: a test call never books a real appointment. You can run the same awkward customer scenario twenty times without putting anything on tomorrow's schedule.
Speech controls most products do not expose
Under Settings, Rex, Speech there are four knobs that matter on real calls. Speaking speed, clamped to a range the realtime API accepts. Silence follow-up seconds, which is how long the agent waits before prompting a customer who has gone quiet. Keyterms, which is a list of words the recognizer should expect, useful for treatment names and neighborhood names; your business name and agent name are included automatically on Grok. And pronunciation overrides, which get injected directly into the realtime prompt, for when the model insists on mispronouncing your surname.
What we do not do
We do not clone voices. The backend route and the storage column exist from an earlier build and the interface is switched off on purpose. Any page that offered you a cloned voice was wrong and has been corrected.
Where the call actually goes
Inbound calls arrive over Telnyx as a media stream. A separate voice bridge service holds the socket and speaks to OpenAI gpt-realtime or xAI Grok on the other side. Text conversations, which is a different pipeline, run on Anthropic Claude.
Go listen
Signup takes about twenty minutes, and the voice picker with real preview audio comes before the payment step. Start at /start, or read how the stack is built.