# Voice & personas

> How the receptionist sounds and behaves: the persona editor, the two voice engines, greetings, languages, and speech behavior.

The persona page is where a phone agent becomes *the business*: its name, its greeting, its voice, and how it behaves on the line.

## The persona

- **Identity** — the name it answers with and a role description ("the front desk at Bright Smile Dental").
- **Greeting** — the exact first line. After-hours can greet differently than daytime.
- **Behavior** — interruption sensitivity, whether it waits for the caller to speak first, what to do with silence, and the voicemail action (take a message, text back, or hang up politely).
- **Transfers** — the staff numbers it may transfer to, by name, with a whisper of context before the human picks up.

## Two voice engines

| Engine | How it works | Choose it when |
| --- | --- | --- |
| Cascaded | Speech-to-text → the LLM → text-to-speech. Full control over the voice model and temperature. | You want maximum answer accuracy and flow control — the default for receptionists. |
| Realtime | A single speech-native model conversing directly. | You want the lowest latency and the most natural interruptions, e.g. the in-widget voice overlay. |

## Testing the voice

The service's test call simulates the whole pipeline from the browser — greeting, flows, transfers — before any real number is attached. Change the persona, call again, hear the difference. When it sounds like the business, go live.

> **VOICE KNOBS ARE PER-AGENT:** Voice, language, and speech behavior belong to the agent, so a draft agent can trial a new voice while the live one keeps answering. Swap when ready — see [Agents](/docs/ai/agents).