Tools, prices and Kukarella's limits checked on 28 October 2026.
A role-play dialogue for staff training is a short scripted conversation, such as an upset customer and an agent, or a manager and an employee, that people hear before they practise the same moment themselves. An AI tool can now write one from a sentence and voice every speaker in about a minute. But a dialogue your staff listen to is a model, not practice. It teaches when you play a poor version and a better one, stop at the moment that matters, ask what the agent should say next, and talk it through afterwards. A meta-analysis of 117 studies of this kind of training found it transferred to the job best when trainees saw both good and bad examples (Taylor, Russ-Eft and Chan, 2005), and a structured debrief adds about a fifth to a quarter to performance on its own (Tannenbaum and Cerasoli, 2013).
We make one of the tools in this guide. Kukarella's DialoguesAI writes the script, casts the speakers and voices them. Your learners cannot answer it back, it has no branches, and the voices it picks for you read every language in one fixed accent, not your staff's: a Quebec French agent, for example, is a voice you choose by hand. Each of those is covered below, with the tools that do what it does not.
In this guide:
- What a recorded role-play is good for, and what it is not.
- Which kind of tool you need: a dialogue you play, an AI that answers back, a video presenter, or no AI at all.
- How to write the request, with a planner that writes it for you.
- How to make it sound like a real call.
- The poor call and the better call, as recordings.
- Running the session: pause, predict, debrief, replay.
- Putting the audio into Articulate Storyline or Rise.
- Sharing, costs, and the questions trainers ask.
What is a role-play dialogue for, if nobody gets to talk?
Short answer: it is the example people hear before they try. A good recorded call shows what the skill sounds like; a poor one shows the mistake your staff actually make. But listening is not practice: after the recording, someone still has to say the words.
Role-play is one of the least-liked things in a training room. In a survey of 281 corporate trainees, 37% named it as the activity they liked least, and it came last of 23 activities for how much people felt it helped them learn (Vampola, 2001). Trainers know the reaction. Adrian Pitt, who runs a training consultancy, wrote that some learners "will sigh and make it quite clear that they HATE HATE HATE role play!" (TrainingZone, 2016). In the comments under his piece, Jamie Lawrence put it down to fear of being judged: "People inevitably think they are being judged on their style, not on the nuances."
A training room where nobody wants to go first.A recording takes the first turn away from the most nervous person in the room. Everyone hears the same call, from the same starting point, before anyone has to perform. That is the part a scripted dialogue does well. It does not need virtual reality or an avatar to count as simulation: David Gaba, a Stanford anaesthesiologist and a pioneer of medical simulation, put it as "Simulation is a technique-not a technology-to replace or amplify real experiences" (Gaba, 2004).
What the research says to do with the example:
- Play a bad version as well as a good one. Taylor and colleagues found transfer to the job "was greatest when mixed (negative and positive) models were presented", and when practice used scenarios the trainees wrote themselves (2005).
- Turn the steps into rules people can say. The same meta-analysis found skill grew most when the learning points were "presented as rule codes": short rules such as "name the problem before you explain the policy".
- Expect the knowledge to fade and the skill to hold. Knowledge from this kind of training decayed over time, while skills and job behaviour "remained stable or even increased".
The "but" is in the last point of that list: it is the practice and the use on the job that keep the skill, not the recording.
Which kind of tool do you need?
Short answer: a dialogue tool if your staff need to hear the call; an AI role-play simulator if they need to say it under pressure; a video avatar tool if the course needs a face on screen; and no AI at all if you have a colleague and twenty minutes. But none of the seven ranking guides we read names a price, so the prices below are from each tool's own page on 28 October 2026.
A dialogue you play (scripted, voiced)
- What it does: writes or takes a script, gives each speaker a voice, and exports audio.
- Examples: Kukarella DialoguesAI writes the script and casts the voices from one sentence (Prime plan $21 a month); ElevenLabs voices a dialogue you write (from $6 a month); Murf and WellSaid are voiceover studios where you write and cast every line yourself (Murf from $29 a month, or $19 billed yearly; WellSaid from $19 a month).
- The catch: the learner listens and cannot answer back.
An AI that answers back (role-play simulator)
- What it does: the learner speaks, the AI plays the customer, and many tools score the attempt.
- Examples: Exec (from $20 a seat a month, billed yearly), Yoodli (from $8 a month billed yearly; team role-plays on business plans with no published price), Second Nature and Mursion (prices on request; Mursion's simulations blend a live person with AI). ChatGPT's voice mode can play an angry customer for free.
- The catch: you control less of what the customer says, and with ChatGPT you get no scores, no tracking, and no say over the data. Never paste a real customer's details into it.
A video presenter (avatar)
- What it does: a talking presenter, and on paid plans branching questions and export to your learning platform.
- Examples: Synthesia (from $19 a month), HeyGen (from $29 a month), Colossyan (from $59 a month on a yearly plan).
- The catch: they are built around a presenter on screen. Colossyan can stage a conversation between avatars, but you still write every line, and video costs more than audio.
No AI
- What it does: a colleague plays the customer from a printed script, or an actor does.
- Examples: the scripts on HubSpot or Tidio, printed for groups of three: one plays the customer, one the agent, one watches.
- The catch: it depends on the people in the room, and nothing is recorded to replay next week.
When something else is the better choice:
- The skill is saying it yourself under pressure, such as calming an angry caller or handling an objection: use a simulator or a live colleague. Hearing a model call is the warm-up, not the exercise.
- You need a score for each learner and proof that they practised: a simulator with a scoring rubric, or the role-play feature of your learning platform if it has one.
- The subject is raw (a bereavement, harassment, a serious injury at work): a human facilitator. isEazy, which sells AI role-play, writes that AI "does not replace the mentor or trainer in high-emotional-complexity scenarios" (2026).
- You already pay for Articulate 360 AI and need one narrator a slide: its own AI voices, which include French (Canada), may be all you need. A dialogue tool earns its place when you need the script written and several speakers voiced in one go.
How do I write the request?
Short answer: say who is talking, what the call is about, the moment it turns, the mistake or the skill you want heard, how long it should be, and the language. But do not expect the tool to add tension you did not ask for: Kukarella's assistant is told to write in a "plain, neutral voice" unless the request says otherwise.
That rule surprises people who expect AI to "make it dramatic". The instructions our assistant runs on name "a customer support call" and "a training role-play" among the dialogues it writes, and then say: "When the user names none of these, write in a plain, neutral voice. Do not add drama, debate or controversy the user did not ask for." So the tension in your call has to be in your request.
A request that works names six things:
- The speakers and their roles. "A client-service agent and a caller", not "two people". Give them ordinary names, and vary them from one scenario to the next (more on that below).
- The situation, with one concrete fact. "His benefit payment is two weeks late and he has called twice."
- The turning point. The line where the call can go well or badly: "The caller says he will go to the newspapers."
- What the agent does. For the poor version, the mistake your staff really make: "The agent quotes the policy before acknowledging the delay." For the better version, the skill: "The agent names the problem, says what she will check, and gives a time."
- The length, in words or turns. "About 220 words" or "twelve turns". A duration also works, since the assistant reckons about 150 words a minute, but it is an instruction, not a guarantee: ask for words if the length matters.
- The language. Type the request in French, or write "in French" in an English request, and the script comes back in French. The voices are a separate step, covered in the next section.
This is what the poor-call request from the planner below looks like in DialoguesAI's request box:
A role-play request pasted into DialoguesAI's request box: speakers, situation, turning point, the mistake and the length.Clark Quinn, a learning-design consultant, puts the rule for the poor version plainly: "These wrong answers, of course, should reflect where learners go wrong, reflecting reliable misconceptions" (Learnlets, 2016). A poor call nobody would make teaches nothing. Ask the supervisors which mistake they hear most often, and write that one.
The planner below turns your answers into two requests, the poor call and the better call, plus where to pause and what to ask.
One more thing the assistant will not do: stage directions. Asked for "[sighs]" or "[angry]", it writes plain spoken lines instead, because its instructions say "Write only what a speaker says." And do not ask for written hesitations ("um", "I... look") hoping they will sound natural: see the next section.
How do I make it sound like a real call?
Short answer: pick a voice for each speaker that your learners would really hear, set the gap between speakers before you first press Play, and listen to every line. But the voices the assistant picks for you read every language in one fixed accent, usually not the one your staff hear, and nothing in Kukarella makes a voice sound like it is on a phone line.
The voices. When the assistant writes a dialogue, it casts the speakers from a fixed set of 22 multilingual voices, by the gender and age it reads in your request. It does not look at accent: each of those voices reads every language in one fixed accent. For an English call that is often fine. For a French call aimed at agents in Quebec, it is not: the script will be in French, read by voices that do not sound Québécois. Kukarella has dozens of native French (Canada) voices, and you choose one for each speaker by hand:
- In the dialogue, open a speaker's card under Actors and click Voice.
- In the Voice Library window, click the accent button and choose French, then Canada. The window now shows French (Canada), with its voices in two tabs: Native voices and Multilingual voices.
- Pick from Native voices, play a few, click Done, then Save & close.
Some voices open only on a paid plan; a locked voice says which plan it needs.
Choosing a native French (Canada) voice for the agent in the Voice Library window.Speed. You can slow a speaker down or speed one up, from half speed to double, in the speaker's window. But the control is missing on some voices: nine of the 22 the assistant casts have none. If you need a slower caller, change the voice first.
Pauses. Real calls have gaps. In the Effects panel, Add pauses between puts a short silence between speakers, from 0.1 to 2 seconds. Set it before the first Play: changing it later voices every line again, and that spends characters again. Three typed full stops ("...") inside a line give a pause of about three-quarters of a second.
Hesitations and emotion. This is where native voices fall short. They give you the right accent and read cleanly, but literally: they carry little emotion, and typed hesitations do not help. When we wrote "Euh… je… bon, écoutez" into a line for a native French (Canada) voice, it came out as a word, a pause, a word, a pause: nothing like a real caller searching for words. The assistant's multilingual voices may handle a hesitation better; we have not tested it, and they read French in their own fixed accent. So choose native voices for the accent, and put the emotion into what people say: "This is my second call, and nobody has told me anything" carries more frustration than any "um".
The phone. To open with a ringing phone, right-click at the start of the first line and choose Upload audio to add your own recording of a ring (or Add sound effect to search Kukarella's library). But the sound plays over the voice, not before it: in our test, the ring and the agent's first words came out together. Right after the sound, right-click again, choose Add pause and set it a little longer than the ring; our 8-second dial-and-ring needed a 10-second pause before the agent started cleanly. There is no filter that makes a voice sound like it is coming down a phone line, and no background noise under the voices. If you need the call-centre hum, add it in your course tool.
Listen to every line. Multilingual voices in particular can slip on a stress or a name. Kukarella's own page says that with them, "After about fifteen seconds most listeners hear a mistake in stress or pronunciation." Fix a wrong line by changing its text, and the next Play voices that line again. Changing a speaker's voice changes every line that speaker has, and each one is voiced and charged again.
Will a synthetic voice put learners off? Early studies found people learned better from a human voice than from a machine voice, in four out of four comparisons (median effect size 0.8) (Mayer, 2005). Later work questions the effect: a 2020 study found "the voice quality did not significantly influence learning", though it did change how much learners trusted the speaker (Chiou, Schroeder and Craig, 2020). Play the call to two or three of your agents before the whole team hears it.
The poor call and the better call
Short answer: make the same call twice, with the same customer and the same facts, and change only what the agent does at the turning point. But keep the poor version believable: if the agent is rude, learners only learn not to be rude, which they knew already.
Here is the pair this guide was written around, made in DialoguesAI on 28 October 2026. The caller's payment is two weeks late; in the first call the agent answers with the policy, in the second she names the delay first.
The poor call: the agent explains the policy before acknowledging the delay.
The better call: the agent names the delay, says what she will check, and gives a time.
And the better call again, written in French from the same request with "in French" in it, then given two native French (Canada) voices by hand. Listen for the accent, and for how evenly the lines are read:
The better call in French, with two native French (Canada) voices.
How we made them:
- One request for the poor call and one for the better call, typed into DialoguesAI's request box. They came from the planner above, in an earlier version that also asked for written hesitations.
- In English, the two voices the assistant cast itself, the same in both calls, so learners hear one change only: the agent's words. In French, two native French (Canada) voices chosen by hand.
- A dial-and-ring recording uploaded at the start of each English call, with a 10-second pause after it.
- Every line played before download, as MP3 and One file.
An older example, made in Kukarella in 2025 for the first version of this page:
A 2025 example: a customer's frustration and its resolution, made in Kukarella.
Vary the customer. The live version of this page named its angry customer "Karen". Do not. A team that hears the same kind of person cast as the difficult one, scenario after scenario, learns the stereotype along with the skill. Change the caller's name, age and voice from one call to the next, and keep the difficulty in the situation, not the person.
The same call, a different caller: change who the customer is from one scenario to the next.How do I run the session?
Short answer: play the call, stop it at the turning point, ask what the agent should say next, then play both versions and debrief. But do not tell people the answer: Cathy Moore, who designs training scenarios, warns that "'Telling' feedback destroys the sense of being in a story" (2011).
A 30- to 40-minute plan with one pair of recordings:
- Set the scene in one sentence. Who is calling, and why. Nothing about what the agent should do.
- Play the poor call to the turning point and stop. Ask: "What would you say now?" Everyone writes one sentence before anyone speaks; that keeps the loudest person from answering for the room.
- Play the rest of the poor call. Ask: "What did the agent do, and what did it cost her?"
- Play the better call. Ask: "What changed, line by line?" Write the answer as a rule people can say, such as "Name the problem before the policy."
- Now practise. Pairs or threes, one customer, one agent, one observer, on a variation of the same call. Swap. This is the part that builds the skill; the recordings only set it up.
- Debrief against the rule. What happened, why, and what you will do on the next real call.
The debrief is not a formality. Across 46 samples, debriefs improved performance "by approximately 25% (d = .67)" over groups without one, and the effect was about the same after a simulated exercise as after real work (Tannenbaum and Cerasoli, 2013).
Replay it later. One session does not hold. Spacing the practice out beats doing it all at once, and the longer you need people to remember, the wider the gap should be (Cepeda and colleagues, 2006). Testing beats re-reading once there is a delay (Roediger and Karpicke, 2006). So a week later, play the first half of the poor call again in a team meeting and ask: "What does she say next?" That takes five minutes.
The need for follow-up is clearest in de-escalation. A randomised trial of de-escalation training at the Louisville Metro Police Department found 28% fewer use-of-force incidents after it (University of Cincinnati, 2022). But the US National Institute of Justice later reported that the training's decision-making model "was the fastest piece to decay", and that "Supervisors were not reinforcing the de-escalation paradigm that patrol officers learned in the trainings" (NIJ, 2023). Police work is not a call centre, but the lesson travels: the supervisors have to hear the same rule.
About the "forgetting curve". You will see charts saying people forget 50% of training within an hour and 90% within a month. The live version of this page had one. We found no study behind those numbers. The curve comes from Hermann Ebbinghaus, who in 1885 measured how much less effort it took him, one person, to relearn lists of nonsense syllables; a 2015 replication with one more subject found the same fast early drop that then levels off (Murre and Dros, 2015). It is a real finding about memory, not a measurement of what your staff forget from a course.
How do I put the audio into Storyline or Rise?
Short answer: download the dialogue By paragraph, which gives one MP3 per line, and place each file on its slide or step. But Kukarella makes the audio only: the branching, the questions and the scoring are built in your course tool.
In DialoguesAI, open Download, choose MP3 under Audio format and By paragraph under Download audio as. You get a ZIP with one file per line, numbered in order from 0 and named with the line's first 20 characters and the speaker, so each line is easy to find. One file gives the whole call as a single track instead, with the pauses you set.
Articulate Storyline 360 takes MP3, WAV, M4A and several other formats; you can drag a file straight onto a slide or use Insert → Audio → Audio from File, and on import Storyline's AI Assistant offers to generate captions (Articulate). For a branching scenario, the per-line files are what you need: put the caller's line on the decision slide, and the agent's poor and better replies on the slides each choice leads to. Rise 360 takes the same files in an audio block.
One of our customers, a public body in Quebec that makes its training in French, works this way: the team downloads the MP3 files and inserts them into their Storyline projects, which they describe as quite easy.
The "but" on Articulate's side. Articulate 360 AI now has its own AI text-to-speech inside Storyline and Rise, with French (Canada) among its languages; it comes with Articulate 360 AI, from $1,449 a seat a year for one person ($1,749 on a team plan) (Articulate). Each script you voice can be up to 5,000 characters, and Articulate's guide describes no way to voice several speakers in one go (Articulate). If you already pay for it and need one narrator, use it. What a dialogue tool saves you is writing the scene and casting several speakers at once.
The text. Download text gives the script as TXT or DOC, but without the speakers' names. For a handout or a transcript, add the names back.
How do I share it with colleagues and learners?
Short answer: a "Can view" link plays without an account; colleagues you invite can listen or edit. But to edit someone else's dialogue, a colleague needs a paid plan or a seat in a paid team.
- For learners or reviewers: open the share icon, set General access to Can view and copy the link. Anyone with the link can play and download the dialogue without an account, once every line has been voiced. A new dialogue starts as Restricted.
- For colleagues: invite up to 10 people by e-mail as editors or viewers; each needs a Kukarella account. One person edits at a time. Since 21 October 2026, an editor needs a paid plan or a seat in a paid team; viewing stays free.
Commercial use. Kukarella includes commercial rights only on paid plans; if the recordings go into your organisation's training, check that this covers your use. Our guide to AI voice commercial use explains what that covers.
What does it cost?
Short answer: writing the dialogue costs nothing in credits; voicing it spends characters from your balance. But the free requests run out: three without an account, ten with a free one.
- Without an account: three requests and 10,000 characters to try it. Nothing is saved.
- Free account: ten requests and 30,000 characters. No commercial use.
- Paid plans: unlimited requests and every voice, but voicing still spends the plan's monthly credits; Prime is $21 a month.
- What voicing spends: the counter next to Play shows an estimate before you press it. A line changed after it was voiced is voiced and charged again.
A 220-word call is about 1,200 characters, so the poor and better versions of one scenario together are about 2,400.
Questions trainers ask
Can learners answer the AI back?
Not in Kukarella: DialoguesAI makes a recording people listen to. For practice where the learner speaks and the AI replies, use a role-play simulator such as Exec, Yoodli or Second Nature, or a colleague.
Can a DialoguesAI dialogue branch?
No. It is one conversation, top to bottom. Make each branch as its own short dialogue, or write both replies as separate lines, then build the branches in Storyline or Rise from the per-line files.
How long should a training role-play be?
Long enough to reach the turning point and show its consequence: one to two minutes, about 150 to 300 words, is enough for most calls. Longer recordings are harder to stop and discuss.
Can I make it in French, or another language?
Yes. Type the request in that language and the script comes back in it. Then choose native voices by hand for the accent your staff will hear; the assistant's own picks read every language in one fixed accent.
Can I use a real customer's call?
Not as it is. Rewrite it: change the names, the numbers and any detail that identifies the person, and keep only the shape of the call. Never paste a customer's personal details into any AI tool your organisation has not approved.
How do I train an AI to hold a complex dialogue?
That is a different job: building or fine-tuning a chatbot. This guide is about scripts that people listen to. If you want staff to practise against an AI that responds, look at the simulators in the tool section above.
Should every scenario have an angry customer?
No. The hardest calls are often calm ones: a refusal, a confused caller, a policy that does not fit the case. Mix them, and vary who the caller is.
How we checked
Kukarella's behaviour was read from the product's code on 28 October 2026 (the assistant's instructions, the casting, the speaker settings, the downloads and the sharing rules), and checked in the product the same day: the requests, the French script, the voice window, the sound at the start of a call and the typed hesitations. The three recordings were made in DialoguesAI on 28 October 2026. Other tools' prices were read on their own pricing pages the same day; prices change, so check before you buy. Every study is linked to its source, and where we could read only the abstract, we quote only the abstract.