Using Call

Choosing and Testing an AI Voice for Your Phone Line

How to pick an AI voice that suits your callers, compare samples, tune speed and pronunciation, give a line its own voice and test it before customers hear it.

Call editorial team · Published · 5 min read

A caller makes up their mind about a phone line within seconds. Before they have heard a single word of content, they have judged the voice: friendly or cold, clear or muddy, hurried or calm. With an AI phone agent, you choose that voice, and the choice matters more than it seems.

A team member wearing headphones comparing audio samples at a desk
Illustrative image

This guide gives you a practical method: how to decide what you need, how to compare candidates, which settings to adjust and how to test. It follows setting up your AI agent and precedes testing the agent with practice calls.

Decide what the voice must do

Before you listen to a single sample, write down what the voice has to achieve. A voice is a tool, and tools should suit the job.

Questions that narrow the choice

  • Who calls?Older callers often prefer a slower, lower and clearer voice. Younger callers tolerate a brisker pace.
  • What language?Pick for the language your callers speak first. A voice that is excellent in one language may be uneven in another.
  • What tone?A clinic wants calm, a restaurant wants warm, a law office wants measured.
  • What is said?Short answers suit a quick voice; long explanations need a steady one.
  • Where do they listen?Cars, shops and noisy streets reward a clear voice with strong consonants.

Your answers give you a short brief: "calm, clear, mid-pace, Japanese or English as needed". A brief makes comparing voices a quick job instead of a matter of taste on the day.

Compare candidates fairly

Voices sound different depending on the sentence, so compare them on the same text. Use real sentences from your business, not a stock greeting.

  1. Narrow the list by language, and by gender if you have a preference.
  2. Pick three or four candidates and play the same sample on each.
  3. Test a greeting, a price-free answer, a number read aloud and a name.
  4. Listen on a phone speaker and through headphones, as callers will.
  5. Ask two or three colleagues to choose blind, without knowing which is which.
What to listen forWhy it matters
Clarity of consonantsCallers on mobiles and in noise need crisp words
Natural pausesRushed phrasing exhausts listeners
Number readingDates, times and phone digits are where voices stumble
NamesYour company and staff names should sound right
ConsistencyThe voice should not change character mid-sentence

Choosing a voice in Call

In Call, voices live under Voice, in AI voices. A banner at the top names the voice in use. You narrow the list by country or language, gender and search, and you can include multilingual voices. Pressing Listen on a card plays a short sample. Pressing Use this voice on the one you want makes it the workspace voice for calls; the card then shows it as currently in use. Only workspace managers see that button, though everyone can listen to samples.

From shortlist to chosen voice

  1. FilterSet the language, gender and any search, and include multilingual voices if you want them.
  2. ListenPlay a sample on several cards.
  3. ChoosePress Use this voice on your pick; the banner updates.
  4. Hear what callers hearPlay the voice with its saved settings, as a call speaks it.
  5. Tune and saveAdjust speed, stability and similarity, try a sentence, and save.

Samples are limited by how many new ones a workspace can make in an hour, and a sample already made does not count again. So decide on a shortlist first rather than clicking through every card.

Tuning speed, stability and pronunciation

Once a voice is chosen, three settings shape how it comes across. Speed controls pace. Stability controls how consistent the delivery is, with lower values sounding more expressive and higher values more even. Similarity controls how close the voice stays to its original character. Change one at a time and listen to your own sentence after each.

Testing like a caller

A voice in a browser tab is not a voice on a phone. Test it in three stages.

Three stages of testing

  • Listen to the samples on the AI voices page
  • Press Hear what callers hear to check the saved settings
  • Type your own sentences in the playground and listen with the current settings
  • Place a real test call to your own phone, on a normal mobile connection
  • Ask a colleague who has never heard it to call and report back
  • Repeat after any change to the voice, the speed or the pronunciation list

If you clone a voice, a test call to your own phone is available for it, which ends after a short time and is marked as a test. Only workspace managers can place it. Cloning needs consent from the person whose voice is used, so agree the permission in writing before you use it. Live voice quality can vary, so judge the voice on real calls, not only on samples.

A short test script

Keep a fixed test script, so every candidate and every later change is judged on the same text. Include the kinds of sentence that expose weak voices.

Play the script on each candidate in one sitting. Listening to them back to back is the fastest way to hear the difference. Write one sentence of notes per voice, such as "clear but a little fast" or "warm, stumbles on numbers", and decide by the notes, not by memory.

Think about every listener

Some of your callers will have hearing difficulties, use hearing aids, or be on a poor line in a noisy place. A slightly slower pace, clear articulation and a short message help all of them. Avoid voices with heavy effects or very low pitch, which can be hard to understand through a small speaker. A hypothetical pharmacy line, where callers may be unwell and anxious, is a good example of a place where clarity should beat personality every time.

Reviewing after launch

After a week of real calls, listen to a handful of them. Do callers ask for repeats? Do they mishear numbers? Do they hang up after the greeting? A change in speed or a pronunciation entry can fix a lot. Review again when you change the agent's script, since a new script may include words the old tuning never tested. Recording prompts and text-to-speech covers the same craft for menu prompts.

Frequently asked questions

Who can change the workspace voice?

Workspace managers can press Use this voice. Everyone can listen to samples.

Can different phone lines use different voices?

Yes. A line can have its own voice, and it keeps that voice when the workspace voice changes.

Why does the voice mispronounce a name?

Voices read names by general rules. Add the name to the pronunciation dictionary with a spelling that makes the voice say it correctly, then listen again.

Is there a limit on samples?

A workspace can make a limited number of new samples an hour, and a sample already made does not count again. Shortlist first.

Should I clone my own voice?

Only if you have the right and consent to use that voice. A cloned voice can be tested by calling your own phone, but live quality should be judged on real calls.

How do I know the voice works on a phone?

Press Hear what callers hear for the saved settings, then place a real test call to your own phone. A voice that sounds fine in a quiet browser may sound different on a mobile line.

Choose and test your AI voice in Call

In the demo console you can try incoming calls in a simulation.

Related guides