Skip to main content
An agent’s voice is a persona name. It is the only thing that selects how the agent speaks — there is no engine, model or provider field in the API, and engine / engine_voice are rejected with a 400 if you send them.
Voice names are case-insensitive on the way in ("keren" works) and always come back Title Case:

Two families

Every voice belongs to exactly one family, and no name appears in both. Naming a voice from the other family is what moves an agent between them — you never say which family you want.

Standard voices

Thirty personas, and the default for every new agent (Rachel). They support the full agent surface: temperature, the three vad_* turn-taking settings, and type: "flow" agents. Michal, Yonatan, David, Rachel, Gil, Noa, Maya, Adam, Shira, Avigail, Amir, Liat, Omer, Tamar, Tom, Benny, Nir, Natan, Yael, Dvora, Yosef, Shir, Anat, Ariel, Roi, Shlomo, Dana, Alon, Ruth, Yuval

Expressive voices

Eight personas with a wider emotional range, more natural interruption handling and their own turn-taking. The accent column applies only when the agent speaks English. All eight speak Hebrew.

What the expressive family does not take

These voices run their own turn-taking and expressiveness, so four settings have nothing to act on. Rather than storing a number that changes nothing about the call, the API rejects them with a 400:
  • temperature
  • vad_stop_secs
  • vad_start_secs
  • vad_confidence
Everything else on the agent behaves identically: system_prompt, greeting_message, agent_speaks_first, language, tools, webhooks, extraction_parameters, silence_timeout_secs, max_continuous_speech_secs, max_call_duration_secs, background_sound and lead_memory_enabled. They are also not available on flow agents (type: "flow"). A flow agent must use a standard voice.

Switching an agent between families

PATCH /agents/{id} with a voice from the other family moves the agent. The move is lossless in both directions — the standard voice is kept aside while the agent is on an expressive one, so switching back restores exactly the voice it had:
If the agent already carries temperature or vad_* values from its time on a standard voice, they stay on the row untouched and are simply not read while it is on an expressive voice. You do not have to clear them to switch — you only cannot set them in the same request.

Availability

Both families are available to every workspace — there is nothing to switch on and nothing to request. All thirty-eight names are accepted on any agent, subject only to the two limits above: an expressive voice takes none of the four turn-taking settings, and cannot be used on a type: "flow" agent.

Errors