Skip to main content
The voice decides how your agent sounds, and also which language and dialect it speaks. This page covers the voice settings, the speech recognition model that hears the caller, and the switcher that lets the agent change dialect during a call.

Where to find it

In the editor’s Configuration panel, open Voice settings. The Language/dialect switcher is in Call settings.
You need permission to read voices to see this section.

Choose the voice

1

Open the voice picker

In Voice settings, click Change on the voice row. Select agent voice opens.
Voice settings in the Configuration panel: the Noura voice with Change, the Expressiveness slider, Voice dictionaries and the STT modelVoice settings in the Configuration panel: the Noura voice with Change, the Expressiveness slider, Voice dictionaries and the STT model
2

Narrow the list

Under Show, pick All, Favorites, In your agents or Project voices. Search by name, or filter by Dialect and Gender. Clear resets the filters.
3

Listen and pick

Press play on a row to hear a sample. Click the voice to use it. The picker closes and the voice is set.
New agents start with Carla, an Arabic voice. To add your own voice from a recording, see Add a voice. For more on browsing voices, see Choose a voice.
Select agent voice with All selected and the Dialect filter open, grouping dialects under Gulf Arabic and Levantine Arabic with a voice count for eachSelect agent voice with All selected and the Dialect filter open, grouping dialects under Gulf Arabic and Levantine Arabic with a voice count for each

Language and dialect come from the voice

There is no separate language or dialect setting. Each voice has a dialect, and the agent takes its language and dialect from it:
  • An Arabic voice makes the agent speak Arabic in that voice’s dialect, for example Jordanian, Saudi or Egyptian. A Fusha voice speaks Modern Standard Arabic.
  • An English voice makes the agent speak English.
  • With the prompt enhancer on, Hamsa adds an instruction to the prompt such as “Always speak in Jordanian dialect of Arabic”, so the model writes in the same dialect the voice speaks.
  • With the enhancer on, the voice’s gender is passed to the model too (“You are a female agent”), so Arabic replies use the right forms for the agent.
To change the dialect, pick a voice in that dialect. The Agents list shows each agent’s dialect.
With Prompt enhancer set to Disabled, no dialect instruction is added. Say in your prompt which language and dialect to speak, or the model may reply in a different one from the voice.

Voice settings reference

There is no voice speed setting.

Speech recognition model

Agents that used the retired Hamsa-STT-S2 now run on Hamsa-STT-S4.
The STT model list open, with Hamsa-STT-S3 (Default), Hamsa-STT-English and Hamsa-STT-S2 (Deprecated)The STT model list open, with Hamsa-STT-S3 (Default), Hamsa-STT-English and Hamsa-STT-S2 (Deprecated)

Voice dictionaries

Dictionaries fix how the agent pronounces names, brands and terms. They belong to the project, so every agent can use them. Pick one or more in Voice dictionaries, or use Manage library and Create a dictionary. When two selected dictionaries have the same entry, the one later in the list wins. See Pronunciation dictionaries.

Let the agent switch dialect during a call

Turn on Language/dialect switcher in Call settings when callers may speak a different dialect, or ask to.
Call settings scrolled to Smart call end and its prompt, with Language/dialect switcher, Interrupt, Ambient sound, Thinking voice and Speaker identification belowCall settings scrolled to Smart call end and its prompt, with Language/dialect switcher, Interrupt, Ambient sound, Thinking voice and Speaker identification below
When the caller speaks or asks for another language or dialect, the model can switch to it. After a switch:
  • The voice keeps its sound and speaks the new dialect.
  • The dialect instruction in the prompt is rewritten for the new dialect.
  • Speech recognition switches between Arabic and English if the language changed.
  • {{ agent_language }} and {{ agent_dialect }} hold the new values.
  • The agent replies in the new dialect right away.
The agent can switch to Saudi, Emirati, Omani, Bahraini, Lebanese, Palestinian, Jordanian, Egyptian, Qatari, Iraqi, Syrian and Kuwaiti Arabic, Modern Standard Arabic, and English.

How it works on a call

  • Everything the agent says goes through the voice, after the voice dictionaries are applied.
  • Everything the caller says goes through the STT model before the model reads it. Recognition can mishear names and numbers, so ask the agent to confirm them (see Write a prompt that works on calls).
  • {{ agent_language }} is ar or en and {{ agent_dialect }} is a code such as jor, uae, msa or en. You can use them in prompts and messages. See System variables.

Good to know

  • Changing the voice changes the agent’s language and dialect at the same time. Check the prompt and greeting still fit.
  • The dialect switcher is offered to the model as an option. It switches when the model decides the caller wants it, not on a keyword.
  • For Arabic agents, User gender detection in Call settings helps the agent address the caller with the right forms. See Call behavior settings.

Choose a voice

Browse, preview and favorite voices.

Pronunciation dictionaries

Fix how names and terms are said.

Write a prompt that works on calls

How the dialect instruction fits your prompt.

Call behavior settings

Gender detection, turn taking and more.