AI role-play by voice or by text: does it matter?

It matters mostly for what you practice: text trains the words, voice trains saying them under pressure. So pick the format that matches where the skill gets used: if your reps will have to perform out loud on a live call, at least the last rounds of practice should be out loud too.

Oleg KulakovCEO, Spellit6 min read

The two formats train different parts of the same answer

Text practice trains the content of what you say. Voice practice trains delivery: starting a sentence without a draft in front of you, recovering after losing your thread, answering while the other side is waiting.

A typed answer can be edited before sending. Nobody on a live call gets that. So a rep who writes a clean reply to a price objection in a chat window has shown that they know the answer, which is useful, but has not shown that they can produce it in the moment before a buyer decides they are being dodged.

That is the operational difference, and it is easy to miss. If you measure readiness by typed practice, you will pass people who know the material and stall when they have to speak it. The failure shows up on the first real call, with a real customer, which is the cost practice was meant to avoid.

Early evidence says voice changes how people talk, and nobody has measured sales results

I found one recent study that compares spoken and typed work with an AI. It is a 2026 preprint, it is not about sales, and it has not been through peer review, so treat it as a signal.

The preprint looked at 957 open-ended debates between university students and an AI opponent (Jain and Wang, submitted March 28, 2026). Spoken exchanges were significantly more verbose and repeated more words and phrases. The authors read the repetition as functional: speakers lean on recurring phrasing to stay coherent while ranging across more ideas. Typed exchanges favored concision and refinement but covered less ground.

So voice changes how people behave in the exchange. This study does not tell us whether they learn more, and it did not measure performance with a real customer, which is the thing a sales leader cares about. If you want the same caution applied to ramp claims, I wrote it up in the piece on whether AI role-play reduces ramp time.

A second early signal comes from medical training, with the same caveat that it is not sales. A 2025 randomized crossover study at two UK medical schools compared one session with an AI voice-based patient against one with human actors. Of 396 students, 378 completed at least one survey. Both formats raised self-rated communication skill, the AI session by less, and the estimated cost per student was about £33 for the AI format against about £62 for actors. It did not compare voice with text, so it says nothing about which format teaches more.

Voice is the only format that tests pace, interruption and silence

Three things cannot be rehearsed in a chat window, and they are the three that tend to go wrong live.

The first is pace. Typing slows everyone to the same speed, so a rep who rushes when nervous looks identical to one who does not. The second is interruption: in text, the buyer waits politely for your whole message. The third is silence. A buyer who goes quiet after a price is information. In a typed exchange there is only an empty box, and nobody feels the pause.

I am reasoning here, not citing a measurement: I did not find a study of these three in sales practice specifically. But they follow from how the formats work, and they give you something to check in any tool you evaluate. Can the customer talk over you? Does it hesitate or go quiet, or does it reply instantly every time? Does it end the call when the conversation goes badly?

Spellit Academy practice is spoken, and a typed line is the fallback

On the Academy page, the role-play is described as a spoken conversation with an AI customer in the browser. The customer opens the call, the rep answers out loud, and the customer replies in a synthesized voice. Reps can talk hands-free, hold a button while they speak, or type a line if speaking is not an option at that moment. The customer can push back, hesitate and go quiet, and can end the call if it goes badly.

When the rep presses Finish, a separate AI model scores the conversation, usually in 20 to 30 seconds. The breakdown quotes the rep's own lines and offers a better phrase for each weak spot, and the rep can run the same scenario again straight away. That quote-and-reason structure is the same idea as a role-play scoring rubric. You can read how the full path works, from modules and tests to the role-play, on the Spellit Academy page, and the sales onboarding use case shows where it sits in a new hire's first weeks.

Match the format to the stage: text to draft, voice to prove

My recommendation, which is a default to start from and not something I have measured: use text where the job is choosing words, and voice where the job is proving you can use them.

The table shows the same default. It is my reasoning from how the formats work, not a measurement.

StageFormatWhat it testsWhy
Learning the answerTextWhether you know what to sayA typed answer can be edited before it is sent
First rounds against pushbackVoicePace, interruption and silenceA chat window cannot reproduce them
Before the first live callVoice, with a pass markSaying it without a draftA live call gives no chance to edit
Speaking is not possibleTextThe content of the answer onlyA rep who is shut out of practice learns nothing

A workable order for a new rep. First, learn the answer and write it down in your own words. Second, say it aloud against a customer who pushes back, several times in the first week. Third, return to it on a schedule, since a spaced repetition rhythm keeps an answer from fading between sessions. Fourth, before the first live call, require a spoken round that clears your pass mark.

Keep typed input for the cases where it is the right tool: an open office, a noisy commute, a rep with a speech or hearing difficulty, a language where the rep is still unsure of spelling. A format that excludes someone from practice is worse than either option.

What to do next

Take the one objection your last three new hires handled worst on a live call. Have each next hire answer it twice: once typed, once spoken to a colleague playing a buyer who interrupts. Note where the two answers differ. If the spoken one falls apart, you have found what text practice was hiding. If you would rather see a spoken round than imagine one, book a demo and bring that objection.

Key points
  • An analysis of 957 debates with an AI found that spoken exchanges ran longer and repeated more, while typed ones were shorter and more polished.
  • Text is the better format for drafting and rewording an answer. Voice is the only format that tests whether you can say it, at pace, with a buyer waiting.
  • Use text early, voice before the first live call, and judge any tool by what happens after the round, not by the format.
See the demo now. Then run it on yours.

A live demo of our products and a real conversation about the growth problem you need to solve.

Book a demo

FAQ

Is voice AI role-play better than text for sales training?

Not in general, on the evidence I found, and the case for voice in sales is narrower. It rests on reasoning: calls are spoken, so delivery under time pressure, interruptions and silence can only be practiced out loud. Use text for working out the answer and voice for proving you can say it. I found no study that followed reps from either format to real customers, so treat claims that voice wins outright as unproven.

Does voice practice cost more than text?

I have no measured ratio I can stand behind, and it depends on the tool. The cost that matters more is rep time: a spoken round needs a quiet place and a few minutes of attention, while a typed round can happen anywhere. If the budget is tight, use text for the first drafts of an answer and keep voice for the rounds that gate a first live call. That puts the more demanding format where it tests something text cannot.

Can reps practice by typing instead of speaking in Spellit Academy?

Yes. The role-play is built around speaking out loud, and the site says reps can talk hands-free, hold a button while they speak, or type a line if speaking is not an option at that moment. The AI customer replies in a synthesized voice either way. I would treat typing as a fallback for the moments when speaking is impossible, not as the way to prepare for a spoken call. The scoring afterward quotes the rep's own lines and suggests a better phrase for each weak spot, and the same scenario can be run again immediately. The trainee also picks the role-play language at the start, so a team that sells in more than one language can practice each in its own.

How many voice role-play rounds does a new rep need?

I have no measured number, so take this as a default to start from: several spoken rounds on each of the three or four objections that come up most, spread over the first week, then a short return visit each week after. Stop counting rounds and set a pass mark instead. A rep who clears it on the first spoken attempt does not need five more, and a rep who misses it after five needs a different kind of help, such as a live call review with a coach. Judge the format by whether the pass mark predicts how the first live calls go.