Playground

Choosing a voice

How a voice for narration differs from a voice for an agent, what to listen for in a preview, and which speech settings matter for each use.

The same library serves two jobs that ask different things of a voice: reading a script you wrote, and holding a live conversation on your behalf. Read this page to pick a voice with the right job in mind, to know what to listen for before you commit, and to know which of the speech settings are worth your time for each.

Narration and agents ask for different things

A narration is a script. You know every word before the voice reads it, you hear the whole result before anyone else does, and you can regenerate until it is right. An agent's voice reads words the agent composes during the call, one turn at a time, to a caller who can interrupt. Nobody hears a draft.

That changes what matters:

Narration on Text to Speech An agent's voice
Who chooses the words You The agent, turn by turn
How long the voice speaks One clip, as long as your text Short turns, for the length of the call
What a mistake costs One regeneration The caller's trust, on a live call
What to optimise for Character and expressiveness for this script Consistency and clarity across many callers
Where you set it The Options panel beside the text The Voice tab of the agent, per voice

A voice that carries a product film need not be the voice that answers the phone. Choose each for its job, and use the same voice for both only when it passes both tests.

The dialect decides first

Before you compare voices, settle the dialect: it is a property of the voice, and it is the first thing a listener notices. A voice in the wrong dialect is the wrong voice however good it sounds. Dialects explains the choice by region; the library's dialect chips narrow the list to one dialect so the rest of this page compares like with like.

What to listen for in a preview

The preview is a fixed sample clip. It tells you the voice's character; it cannot tell you how the voice reads your words. Use it to shortlist, then test on your own text.

  • Pace. The preview is the voice at its natural pace. A voice that sounds hurried in its own clip will sound hurried on your script; prefer a voice whose natural pace fits, over one you would have to slow down.
  • Clarity at the end of phrases. Drag the seek slider to the end of the clip and listen to how the voice finishes a sentence. That is what a caller hears just before they reply.
  • Warmth against authority. A support agent needs the caller to feel heard; an announcement needs to be believed. The row's tone and pace give a hint, but decide by ear.
  • Fit with the other voices on the agent. When an agent carries several dialects, play the candidates back to back and keep the set that sounds like one brand.

Then press Use in Text to Speech. Text to Speech opens with that voice selected: paste a real paragraph from your script, or a typical agent reply, and generate it. Generating is charged in credits; the preview and the library are not.

The settings that matter for narration

On Text to Speech, the Options panel holds three controls for the selected voice, and the writing area adds a fourth for an Arabic voice.

Control Range Default Worth your time when
Speed 0.5 to 2.0, in steps of 0.05 1.0 The read is right but the pace is not. Slow down for instructions and legal text; speed up for energy
Guidance 1.0 to 5.0, in steps of 0.1 2.0 You are comparing renders of the same text and want to test the guidance scale. Leave it at the default otherwise
Quality 16, 32, 48, 64 or 128 48 You are comparing renders of the same text. Leave it at 48 otherwise
Tashkeel Arabic voices only Off The text is Arabic and a word is being read with the wrong vowels. Mark the words you care about yourself; they stay as you typed them, and the rest are added for you

Change one at a time and regenerate; a change to two controls at once tells you nothing about either. Reset to defaults puts the options back on their defaults. Voice settings describes every control.

The settings that matter for an agent

On the agent's Voice tab, each voice under Used Voices has its own Voice Settings, opened from the card. The sheet holds cfg, speed and num_steps, plus Tashkeel (ON or OFF) on an Arabic voice. On the agent, speed and cfg move in steps of 0.5; the ranges and defaults otherwise match Text to Speech.

For an agent, the settings that matter are the ones you set once and stop touching:

  • The default voice. Press Make default on the card that should open every conversation. Exactly one voice is the default.
  • One voice per dialect the agent serves, so that dialect detection has a matching voice to switch to. See Voices and language.
  • speed, only if callers say the agent talks too fast or too slowly. Move it one step and listen to a real call before moving it again.
  • Tashkeel ON on an Arabic voice that reads Modern Standard Arabic, so the vowels are right on every turn without anyone marking the text.

Leave cfg and num_steps at their defaults on an agent unless you have run your own listening tests. A change applies from the agent's next conversation; a call already running keeps the settings it started with.

Get it right

Choose by listening to your own words, then change one setting at a time. A voice picked from its name and tuned on four sliders at once is a voice you cannot get back to.

Recommended

  • Narrow to the dialect, play the previews, save two candidates, press Use in Text to Speech, generate the same paragraph with each, and then move Speed alone if the read needs it.

Less effective

  • Pick the first voice whose name you like, set Speed, Guidance and Quality in one go, and judge the result on a single sentence.

Worked example

Acme needs a voice for a two-minute onboarding video and a voice for its support agent, both in Egyptian Arabic.

For the video, the team presses the Egyptian chip, plays each preview and saves two warm voices. They press Use in Text to Speech on the first, paste the opening paragraph of the script and generate. It is right but too quick, so they move Speed from 1.0 to 0.5 and generate once more. That clip is the one they download.

For the agent, they want the calmer of the two. They open the voice's details panel, press Add to an agent, choose Acme Support and press Add voice. On the agent's Voice tab they press Make default on the new card and leave Voice Settings untouched. Their first test call sounds like the preview did, and that is the point: no setting to remember, nothing to drift.

FAQ

Can the same voice serve a narration and an agent?

Yes. A voice is not reserved for either use. Test it on both a script and a typical agent reply before you decide, because the two reads differ.

Do the settings I choose on Text to Speech carry over to the agent?

No. Text to Speech settings apply to what you generate there. An agent's voice has its own Voice Settings on the agent, and each agent tunes its copy of the voice separately.

Do I need a Fusha voice for diacritics?

No. The Tashkeel switch on Text to Speech and Tashkeel on an agent are offered on any Arabic voice. English voices do not have them.

How do I hear a voice on my own text before I commit to it?

Press Use in Text to Speech on the row, then generate your text on Text to Speech. Generating is charged in credits.

Does changing a voice's settings cost anything?

No. Changing a setting is free. Credits are charged when speech is generated or when an agent speaks on a call.