Multi-Speaker AI Dialogue: Which Tool, and How

Tools, limits and prices checked on 1 October 2026.

To make an AI dialogue that sounds like people talking, you need four things: a script written for the ear, a clearly different voice for each speaker, the right pauses, and a tool that voices the whole conversation. Two kinds of tool do that last part. Dialogue models voice the conversation in one pass: on 1 October 2026, Google's Gemini TTS takes two speakers per request and ElevenLabs' Text to Dialogue up to ten. Studio apps voice it line by line, and Kukarella's DialoguesAI also writes the script and gives every speaker a voice. But no tool knows which accent your listeners need. That choice is yours, and it matters.

It matters because listeners understand best the voices they have practised with. In one study, native English listeners who had practised with five Chinese-accented speakers understood a new one better than listeners who had practised with only one (Bradlow & Bent, 2008). And the tests your learners sit already mix accents: IELTS says that "different accents, including British, Australian, New Zealand and North American, are used" in its listening test (IELTS). A dialogue with two speakers in two accents is a quick, inexpensive way to give learners more than one voice — if you set it up well.

We make one of these tools, DialoguesAI. This guide names other tools where they fit better, and it says what ours does not do.

The method in five steps:

1. Write it for the ear: short turns, contractions, reactions, one moment where someone checks a fact.
2. Give every speaker a voice that sounds clearly different, in the accent your listeners will meet.
3. Set the pauses your listeners need: close to real conversation for advanced listeners, longer for beginners.
4. Generate it, then listen to every line before anyone else hears it.
5. Export it the way you will use it: one file, or one file per line.

Why a dialogue, and why more than one voice?

Short answer: people often learn more from overhearing a good conversation than from the same lesson told as a monologue, and listeners who practise with several voices often cope better with a voice they have never heard. But neither happens by itself. The dialogue has to ask the questions a learner would ask, and the learners have to do something with it.

A direct test of the format comes from a team at Arizona State University. Five tutors taught the same science lesson in two ways: to a student on camera, as a dialogue, and alone to the camera, as a lecture. Students then watched one of the two videos in pairs, with a workbook. Those who watched the dialogue learned as much as the students who had been tutored in person. Those who watched the lecture made no significant gain (Chi, Kang & Yaghmourian, 2017). The authors' explanation: the dialogue made the watchers ask and answer questions themselves.

The catch is in the conditions. In an earlier study led by the same researcher, Michelene Chi, at the University of Pittsburgh, students who watched a tutoring session together, talking it through, learned about as much as the student being tutored; students who watched it alone made no significant gain (Chi, Roy & Hausmann, 2008). A dialogue is not magic. A dialogue with a task attached is a good lesson.

What makes a dialogue teach: a speaker who asks the questions a learner would ask, and a speaker who makes the mistake a learner would make. In a study of 364 first-year physics students in Sydney, the two versions of a lesson that said common misconceptions out loud — a student–tutor dialogue, and a lecture that stated and refuted them — produced much larger gains than a plain explanation (Muller and colleagues, 2008). The active ingredient was the mistake said aloud; a dialogue is the natural place to say it. The study's author, Derek Muller, went on to start the YouTube science channel Veritasium.

Why more than one voice. One voice teaches the ear to understand one voice. In Bradlow and Bent's study, native English listeners who practised with one Chinese-accented speaker scored about 82 on a new Chinese-accented speaker. That was no better than practising with native speakers. Listeners who practised with five speakers of that accent scored about 90, as well as if they had practised with the new speaker himself (Bradlow & Bent, 2008). A follow-up study trained listeners on five speakers with five different first languages. Afterwards they understood a Slovak-accented speaker better, although they had never heard that accent. Listeners trained on one accent did not (Baese-Berk, Bradlow & Wright, 2013).

The "but": the evidence is mixed. A larger 2022 study with 166 Japanese learners of English, trained to hear English r and l, found no difference in how well they learned from many voices or few, and only ambiguous evidence of a benefit with new voices (Brekelmans and colleagues, 2022). A review of 18 studies found a clearer benefit with new voices, but with wide uncertainty (Zhang, Cheng & Zhang, 2021). Several voices are a good bet for the voices your learners will meet outside the classroom, not a guarantee.

The tests already require it. Cambridge tells teachers that listening practice "should include a range of voices, accents and styles of delivery" (B2 First handbook). Pearson says PTE Academic test takers "will hear international English speakers with a variety of accents" (PTE handbook). And when researchers played the same TOEFL lecture to 21,726 test takers in US, British and Australian voices, they found that accent strength and familiarity affect comprehension "even with quite light accents" (Ockey & French, 2016).

If you train adults on a job, not on a language, turn this around. In a lesson that is not about listening, an unfamiliar accent is extra effort: students found a foreign-accented narrator harder to understand, and did worse at applying the lesson (Chan and colleagues, 2020). Cast the accents of the customers and colleagues your trainees will really talk to, and give them a transcript: in the same study, full on-screen text helped learners catch the words, though not apply them.

Which AI tools can make a dialogue with several voices?

Short answer: two kinds. Dialogue models voice the whole conversation in one pass, so the turns flow into each other — but they limit the number of speakers or the length of a request. Studio apps voice each line separately, so you can change any line or any voice — but the speakers never overlap, and the gaps are the ones you set. Microsoft puts the difference plainly: single-speaker voices "synthesize each turn in isolation", while its multi-talker voices keep the whole exchange together (Microsoft Learn).

The old way: each voice made separately, then stitched together.

Before dialogue models, every speaker was generated separately and the clips were joined in an audio editor. Teachers still do it. The teacher trainer Ross Thorburn calls the change "the biggest update": "Teachers no longer always need to create one voice at a time and edit the files together" (Hong Kong TESOL, June 2026).

The main options, checked on 1 October 2026:

Google Gemini TTS

Google's speech models in the Gemini API take "up to 2 speakers using prebuilt voices" in one request, and Google says that custom or cloned voices in a dialogue must be generated one speaker's turn at a time (Gemini API docs, updated 30 September 2026). Google's newest model, Gemini 3.8 Flash TTS, became generally available on 22 September 2026; Google's announcement promises "2,000+ production-ready voices", including regional varieties such as Mexican Spanish and Quebec French (Google), while Google documents its two-speaker mode only for its prebuilt voices, which it lists as 30 studio voices.

Best for: developers who need two speakers in one natural-sounding take, at a low price per minute. The but: it is built for developers — an API, with AI Studio to test it, not an app for teachers; a third speaker means a second request; and you choose each voice yourself.

ElevenLabs

ElevenLabs' Text to Dialogue turns a list of speaker turns into one audio file, on its Eleven v4 and v3 models, which ElevenLabs says cover more than 90 and more than 70 languages (ElevenLabs docs). Its API takes up to 10 different voices per request (API reference). Inline tags such as [laughs] or [whispers] steer the delivery, and punctuation can mark an interruption.

Best for: expressive scenes with laughter, sighs and interruptions. The but, in ElevenLabs' own words: keep a request "at or below 2,000 characters" for reliable results, "the models are nondeterministic", "several generations might be required", and the audio tags are "still under active development". Its free plan carries no commercial rights (ElevenLabs docs).

Microsoft Azure Speech

Microsoft's multi-talker voices speak a conversation written in its markup language, one turn after another, in a single voice model (Microsoft Learn). Since March 2026 they cover French, Spanish, German, Italian, Brazilian Portuguese, Korean, Japanese and Chinese as well as English, and the English model has eight named speakers (Microsoft's release notes). They are high-definition voices, which Microsoft still lists as a preview.

Best for: developers already building on Azure. The but: you get the speakers Microsoft built into each model, not any voice you like, and the timed-pause and speed controls of Azure's standard voices do not apply to them.

Gemini Notebook (formerly NotebookLM)

Google's Audio Overviews made two-voice AI audio famous in 2024: upload your documents and two AI hosts discuss them (Google, September 2024). Google renamed NotebookLM Gemini Notebook on 16 July 2026 (Google). Overviews come in 80+ languages and four formats — Deep Dive, Brief, Critique and Debate — and you can steer the topic and the expertise level with a prompt (help page).

Best for: a spoken summary of your own sources, in minutes. The but: you do not write the lines, choose the voices or pick the accents, and Google warns that overviews "may contain inaccuracies or audio glitches". It is a summary tool, not a dialogue tool.

Studio apps: Murf, WellSaid, Typecast, Narakeet and others

These apps voice a script block by block, with a voice per block. Murf's help centre says it plainly: put "each character's dialogue in distinct blocks" and pick each block's voice (Murf). Narakeet lets you name the voice inside the script (Narakeet). Prices differ widely: on 1 October 2026 Typecast's cheapest plan with a commercial licence was $5 a month (Typecast), and every WellSaid plan below Enterprise offered English voices only (WellSaid).

Best for: narration and dialogue that you edit line by line. The but: the turns never overlap, so the result sounds like people taking turns, not like people interrupting each other.

Open-source models, and the tools that disappeared

Dia, from the two-person start-up Nari Labs, generates a two-speaker dialogue from a script marked [S1] and [S2], with sounds such as (laughs) (GitHub). Its successor Dia2 is Apache-licensed but, in Nari Labs' words, "only supports up to 2 minutes of generation in English" (GitHub). Dia needs a graphics card for now, and Dia2 can fall back to an ordinary processor; both have a free online demo.

Three warnings before you build a course on a tool. Microsoft's open-source VibeVoice could voice up to 90 minutes with four speakers, but eleven days after release Microsoft wrote that it had found "instances where the tool was used in ways inconsistent with the stated intent" and removed the code (GitHub). PlayHT, still recommended on many "best of" lists, was bought by Meta in July 2025 under its newer name, PlayAI (TechCrunch), and on 1 October 2026 its website no longer loaded. Keep your scripts and your exported audio: tools change faster than courses.

Which one sounds most natural?

No major public leaderboard ranks dialogue today. On the best-known one, where people vote blind between two voices reading the same text, ElevenLabs' Eleven v4 ranked first on 1 October 2026 (Artificial Analysis). But the votes compare single voices, not conversations, and the accent filter covers only US and UK English. Test with a page of your own script.

How to create a multi-speaker dialogue in Kukarella

Short answer: describe the conversation in one sentence, check who the assistant cast, give each speaker the accent you need, set the pauses, press Play, then export. The assistant writes the script and casts the speakers in seconds. But it never chooses an accent, so the third step is yours, and it is the one that matters most for learners.

Who decides what. Every dialogue rests on four decisions. The assistant makes the first two in seconds; the accent and the pauses are yours:

The assistant writes the script and picks the voices; the accent and the pauses are yours.

Watch it done in one take: the request the planner below writes by default, the script and the cast in under half a minute, two recasts in the Voice Library, the pauses and the download. Five minutes, recorded on 2 October 2026, unedited and without narration; what you hear are the voice samples and the dialogue. The script and the cast arrive at 0:27, the recasting starts at 1:10, the pauses at 4:25.

Step 1: describe the conversation

Open DialoguesAI, click Use Assistant and type what you need. You can start from the home page as well, without an account.

Kukarella's home page: describe a dialogue, no account needed.

A good request names the language, the level or exam, each speaker with gender and age, the situation and the length. For example: "Write a 2-minute dialogue in English for language learners at CEFR level B1. Situation: buying a train ticket at a station. Speakers: Ana (a woman, about 30, from the US); Tom (a man, about 55, from the UK)." The planner further down writes this request for you.

The assistant's panel in DialoguesAI, where you type the request.

What the assistant does with it (as of 1 October 2026):

• It writes to the level you name: CEFR A1 to C2, HSK, JLPT or TOPIK, or the format of an exam such as IELTS, TOEFL iBT, TOEIC, DELE, DELF or Goethe.
• It writes in the language you name, or in the language of your request. An attached file — PDF, DOCX, TXT, JPG or PNG — gives it the content, never the language.
• It aims at about 150 words for each minute you ask for.
• Writing the script costs no credits. A guest can ask for three scripts and a free account for ten, in total; paid plans have no limit.

The but: the level and the length are instructions to the AI writer, not guarantees. Our request for a two-minute dialogue came back at about 700 characters, which plays in under a minute, and nothing said it was shorter. Read the script and check its length before you go on.

Step 2: check who the assistant cast

In under half a minute you get a project: the script, one card per speaker, and every line already assigned to a speaker. Nothing is voiced until you press Play.

A finished project: one card per speaker, every line assigned.

The assistant gives each speaker a voice by gender and age, from a short list of voices it uses for every script. That is why the request should say each speaker's age: the assistant reads your request first. But it does not choose accents. A speaker you describe as British, Indian or Nigerian gets a voice from that same short list, not a voice from that country.

Step 3: give each speaker the accent

Click a speaker's card, then Voice. Choose the language and the country — English (India), English (UK), Spanish (Mexico) — listen to two or three voices, pick one, click Done, then Save & close. Every line of that speaker changes to the new voice. But each country's list also holds multilingual voices, the assistant's own among them, so listen for the accent; once you have saved, the speaker's card should show that country's flag.

Changing a speaker's voice: the speaker's card and the Voice Library.

Hear it — one dialogue the assistant wrote from the planner's default request. The request made Tom British; the assistant gave both speakers voices marked as US English. Then the same dialogue with two pairs of native voices:

Kukarella has 2,000+ voices. Its Voice Library lists English in 14 regional varieties, Spanish from 22 countries and French from 4. Two buts. Some varieties have only a couple of native voices — the rest of that country's list is multilingual — so check before you write a scene with three speakers from the same country. And some voices open only on a paid plan; every voice plays a sample first, so you can listen before you decide.

Native or multilingual? A native voice speaks one language, in the accent of its country. A multilingual voice speaks many languages, often with the same accent in all of them. For listening practice in an accent, choose a native voice. For a speaker who switches languages in the middle of a scene, choose a multilingual one.

Step 4: set the pauses

Open Effects. Under Add pauses between, switch on paragraphs: it puts a pause at the end of every line — in a dialogue, the gap between two speakers. It is off by default and runs from 0.1 to 2 seconds. You can also click anywhere in a line and add a pause of half a second to 90 seconds, and three full stops typed one by one (...) make a pause of about three quarters of a second; the single "…" character some keyboards put in makes none. How long the gaps should be for your listeners is in the section on pauses below; the planner suggests them by level.

The Effects panel, with pauses between sentences and paragraphs.

The but: pauses, speed, pitch and volume work on most voices, not on all of them, and volume needs a paid plan, so listen after you change them.

Sounds help too. Type / in a line to add one of about 2,900 sound effects, or your own MP3, WAV or OGG file. Pauses and sounds cost no credits themselves. But adding one to a line that already has audio voices that line again, and charges it again — and switching on the paragraph pauses after your first Play does that to every line. Set them first.

Step 5: press Play, and listen to every line

Play voices every line that has no audio yet; when it has finished, press Play again to hear the conversation. Our test dialogue, under a minute long, took 12 seconds to voice on 2 October 2026; a dialogue of that size takes about 10 to 15 seconds, depending on the voices. Change a line afterwards, and only that line is voiced again and charged. Give a speaker a new voice, and all of that speaker's lines are.

Listen to all of it before your learners do. When Duolingo generated listening conversations for its English test and voiced them with AI, reviewers found that 477 of 3,245 voiced turns — 15% — needed fixing for pronunciation, stress or intonation (Runge and colleagues, 2024). Give names, numbers and acronyms a second listen: how they are read varies from voice to voice. If a voice reads one wrongly, change the spelling or try another voice.

Step 6: export it

Click Download and choose:

• By paragraph — a ZIP with one numbered file per line, each named with its speaker. Use it for slides, a course tool or a video timeline.
• One file — the whole conversation. The gaps between speakers are the pauses you set; with no pauses, the lines follow each other directly.
• Text, as TXT or DOC. The text download leaves out the speakers' names, and so does copying the text from the page, so add the names yourself; the By paragraph file names carry them.

The download box: one file, one file per line, or the text.

There are no subtitle files in DialoguesAI. Publish a transcript that names the speaker of every line. The web's accessibility standard asks for a text alternative to any recorded audio-only content, and W3C's guidance says a transcript "identifies who is speaking" (W3C). For a listening task, hand it out after the task, as Cambridge advises, or it becomes a reading task.

To share it, set General access to Can view (a new dialogue starts as Restricted): anyone with the link can then play it without an account. Or invite up to 10 people who already have a Kukarella account; one person edits at a time.

Pictures for slides or a video

Create Visuals makes images and short clips from your script. The Ideas button suggests pictures once the script is longer than 200 characters. An image costs 2 AI Art & Video credits and a 10-second clip 20; Prime includes 40 such credits a month (Pricing). The but: the clips are silent and wide (16:9). They illustrate a dialogue; they do not show people speaking it.

Create Visuals suggests pictures from your script.

Pictures made with Create Visuals.

If you would rather write it yourself, or translate it

Write it yourself: click Create Project, type the lines, create one speaker with Create actor for each person, and assign each line to a speaker. The but: pasting a script with NAME: at the start of each line does not create the speakers; the names stay in the text and would be read aloud, so delete them after you assign the lines. A project holds up to 500 lines.

Translate it: there is no one-click "translate and re-voice". Once every line has been voiced, Ask AI on a line offers Translate to on a computer, and the translated line keeps its speaker and voice. But a guest can use Ask AI three times and a free account ten, in total, and after you replace one line the Ask AI buttons come back only when you press Play again. Or paste the script into a new request to the assistant and name the other language: it writes a new script, casts it again, and uses one of your script requests. Either way, give each speaker a native voice for the new language, as in step 3.

What it costs

Writing the script costs nothing. Voicing costs one credit per spoken character in any ready-made voice; speaker names, pauses and sound effects are free. A two-minute dialogue is about 1,800 characters. Prime costs $21 a month for 300,000 credits and includes commercial use; the free plan does not (Pricing, 1 October 2026). Kukarella has no public API, so if you need dialogue inside your own app, use one of the APIs above.

How to write a dialogue that sounds like real people

Short answer: write short turns with contractions and reactions, build it like a real exchange — an opening, a check, a closing — and give the hesitations to the speaker who is unsure. But keep it cleaner for beginners: scripted, tidy speech is easier to understand, and the realism can grow with the level.

The people who built Google's Audio Overviews found the same. On the Hard Fork podcast in September 2024, NotebookLM's editorial director Steven Johnson explained that the system turns a dry script into one with banter and pauses, because two robots talking to each other are unbearable to listen to (as transcribed by Simon Willison). And in a 2025 test by the voice company Sesame, listeners showed no clear preference between its AI lines and real recordings heard one by one — but heard inside a conversation, they "consistently favor the original recordings" (Sesame). A line can sound human on its own. In a conversation, its delivery gives the AI away.

What real conversation looks like, in numbers:

• Turns are short: about 8 to 10 words per turn in American telephone calls (Yuan, Liberman & Cieri, 2006).
• Listeners talk back: short reactions such as uh-huh, yeah and right are about one in five of all utterances in such calls (Stolcke and colleagues, 2000).
• People stumble: about 6 times in every 100 words — fillers, repeats and restarts (Bortfeld and colleagues, 2001).

Older textbook dialogues left most of this out. When the linguist Alex Gilmore compared coursebook conversations with real ones, the textbook versions ran as tidy question-and-answer pairs, while the real ones had false starts, repetition, overlaps, hesitations and back-channels — and the person giving information kept asking for clarification (Gilmore, 2004).

The but: tidy is easier to follow. Intermediate learners of Spanish scored significantly higher on scripted recordings than on unscripted ones of the same content (Wagner & Toth, 2014). So use clean dialogue for beginners and first listenings, and add the stumbles, reactions and natural speed as learners advance.

Rules that work with AI voices:

• Keep most turns short — one or two sentences. Let a long explanation be interrupted by a question.
• Write spoken grammar: I'm, don't, Coming back Sunday, Right. Sure.
• Build it like a real exchange: an opening, at least one check (Sorry — platform four?), and a closing.
• Repeat key facts — a time, a price, a place — the way real people confirm them. For a listening task, do what exam writers do and let a wrong number come first: Cambridge's A2 sample test mentions "July the eighteenth" and "the eleventh" before the answer, "the twenty-fifth" (A2 Key handbook).
• Start a turn with a word or two before the key fact — Right, so…, It's…. Every change of voice costs a listener a moment, and that cost falls by more than half once the new voice has been heard for about 0.6 seconds (Choi & Perrachione, 2019).
• Give one speaker the learner's question or mistake, and let the other answer it.
• Give the hesitation to the speaker who is unsure. In a study of synthetic speech, a filler such as uh in the middle of a sentence did more than pitch or speed to make the speaker sound unsure (Kirkland and colleagues, 2022). In a role-play, the customer may hesitate; the agent should not.
• Write no stage directions. AI voices read them aloud: ElevenLabs warns that its model "will still speak out" emotional cues written as prose (ElevenLabs). Put the feeling in the words.
• Read it aloud yourself before you generate it. If you stumble, rewrite the line.

An example, written to these rules — a B1 listening dialogue we wrote for this article. Introduce it the way exams do: "You will hear Ana, a visitor from the United States, buying a train ticket from Tom at a station in London."

The hesitation belongs to the traveller, not the clerk. The real price comes after a wrong one, and Ana checks it. The pause sits inside Tom's turn, after the platform, so it sounds like someone letting a fact land, not like a slow reply. Write the pause with your tool's pause control, not as a word: an AI voice would probably read "(pause)" aloud.

How to write pauses, interruptions and laughter

Short answer: punctuation is the one mark every engine reads. For a timed pause, use the tool's own pause control; for laughter, its own tag or a sound. Real interruptions happen only in dialogue models — line-by-line tools play one turn after another.

How long should the gap between speakers be? In real conversation, almost none. Across ten languages, answers came on average about a fifth of a second after the question ended — from 7 milliseconds in Japanese to 469 in Danish (Stivers and colleagues, 2009). Slow answers carry a meaning: in the same data, a no came on average 100 to 500 milliseconds later than a yes — significantly so in seven of the ten languages — and non-answers such as I don't know came later than answers (Stivers and colleagues, 2009). So a dialogue with a long silence after every line does more than sound read: it can make a yes sound reluctant.

The but: learners need time. The Council of Europe's levels describe an A2 listener who can follow "short, simple social exchanges, conducted very slowly and clearly", and only at C1 a listener who follows overlapping talk "at a natural speed" (CEFR Companion Volume, 2020). Pauses, hesitations and repetition give that time. In Eileen Blau's studies with beginner learners in Puerto Rico and Japan, short talks with pauses — especially pauses filled with well, I mean or uh — were understood better than the same talks read straight through (Blau, 1991). A moderately slower voice may help lower levels too, but ElevenLabs warns that extreme speed settings cost quality. So keep the gaps between speakers short, put a pause inside a turn after a key fact, and give thinking time outside the dialogue — before and after it, the way exams do.

Starting points — our synthesis of the studies above, not a finding of any of them. Listen and adjust:

In Kukarella, the paragraph pause starts at about 0.3 seconds when you switch it on. Listen before you raise it.

The sources for the other tools are their own pages: the Gemini API, Google Cloud, ElevenLabs, ElevenLabs prompting, Microsoft and Azure HD voices, all read on 1 October 2026. The but for all of them: tags and marks change how a line sounds on some voices more than on others. Listen to every tagged line.

Check yourself. Three quick questions on what you have read so far:

Which voices and accents should you cast?

Short answer: cast for contrast — a different sex, age or accent for each speaker — and cast the accents your listeners will meet. But check that the accents exist before you write the scene, and keep the cast small for beginners.

Contrast first. With only their ears to go on, listeners tell voices apart mostly by pitch and by the resonance of the voice, the same cues from which they hear a speaker's sex, size and rough age (Badcock & Chhabra, 2013). Cambridge's sample tests follow the same pattern. In the official sample tests for A2 Key and B1 Preliminary, every one of the 21 recordings with two speakers pairs a man and a woman (our count of the A2 Key and B1 Preliminary handbooks). In the longer parts they also introduce the speakers before anyone talks, often by name: "You will hear Robert talking to his friend, Laura, about a trip to Dublin." A BBC drama producer's guide to radio explains why: the listener only knows a character exists "if that character speaks, or if another character refers to him or her by name" (Fiona Ledger, How to Write a Radio Play).

Keep the cast small. Cambridge's A2 and B1 recordings never have more than two speakers, and the same BBC guide warns: "Do not have more than 6 characters in a half hour play. There is a risk of confusion if you do." Our rule of thumb: two voices up to B1; three or four for advanced listeners, if each one is clearly different.

Then the accents your listeners will meet. For exam practice, follow the exam: IELTS names British, Australian, New Zealand and North American accents. For work, think beyond them. When researchers voiced a mock TOEFL listening test with speakers of six varieties of English — American, British, Indian, South African, Chinese and Mexican — test takers did not score significantly lower on the non-native speakers who were highly comprehensible than on the native ones (Kang, Thomson & Moran, 2018). For many learners, much of the English they hear at work will come from other non-native speakers.

Choosing one voice — by language and accent, native or multilingual, man or woman — has its own guide: How to choose the best AI voice. For classroom listening dialogues in depth, see AI for language teachers.

Plan your own dialogue

Answer five questions and the planner below writes the request for the assistant, lists the accent to choose for each speaker, gives the pauses to set for your listeners, and works out the size of the job. It counts credits the way Kukarella charges them: one per spoken character.

Is an AI dialogue good enough for teaching and training?

Short answer: for practice dialogues, role-plays and course updates, usually yes. In two studies of narrated lessons, learners learned about as well from a modern AI voice as from a recorded human one; none of the studies tested a dialogue. But not every study agrees, and AI voices still slip on stress and intonation, so listen to every line and add real recordings too.

In a 2019 study by Scotty Craig and Noah Schroeder, learners who heard a modern synthetic voice were not statistically different in what they learned from those who heard a recorded human (Craig & Schroeder, 2019). With 51 engineering students learning in English at about B1 level, a human narrator scored 60.6 on the test, a modern AI voice 54.1 and an old-style synthetic voice 42.7; the difference between the human and the modern voice was not statistically significant (Dinçer, 2022). But a small 2024 study of 30 learners found that the recorded human voice "significantly improved learning performance" (Jing and colleagues, 2024).

Teacher trainers say the same in plainer words. Ross Thorburn warns that "multi-speaker tools may mix up speakers, create odd pauses, make emotions too strong, or produce dialogue that sounds too clean", and advises teachers to supplement AI audio with authentic listening texts (Hong Kong TESOL). Nik Peachey, writing for the British Council's TeachingEnglish site, says anything we produce with AI "should be carefully checked and edited" (TeachingEnglish).

When something else is better:

• The performance is the point — a drama for an audience, a scene that must move people: hire actors.
• Learners must answer back — a role-play is often about the learner's own reply: record your colleagues, or have learners act it out. An AI dialogue is for listening.
• There is no exchange to follow — a procedure, a policy, a phone menu: one clear voice is easier to follow and quicker to update.
• You need it inside your own app, or with faces that speak — use an API or a video avatar tool.

Tell your listeners the voices are AI. OpenAI, for one, requires "a clear disclosure to end users that the TTS voice they are hearing is AI-generated" from anyone who uses its voices (OpenAI). It is good practice whatever tool you use.

Whose voices can you use?

Short answer: voices from a tool's library, playing characters you made up, need nobody's consent. A real person's cloned voice needs that person's permission — get it in writing. And conversations with several familiar voices are already a tool of fraud.

In early 2024 an employee of the engineering firm Arup joined a video call with what looked and sounded like the company's finance director and other colleagues. They were deepfakes. The employee made 15 transfers, about US$25.6 million in total (Fortune, May 2024). Several familiar voices in one conversation are what made it believable (CNN, February 2024).

The big platforms now ask for consent before they clone anyone. Microsoft requires "explicit written permission from voice talent" for a custom voice (Microsoft Learn). ElevenLabs lets you make a professional clone only of your own voice (ElevenLabs). In Kukarella, a voice you cloned appears in the Voice Library and can play a speaker like any other voice; clone only your own voice, or one whose owner has agreed.

Questions people ask

Which text-to-speech tools support multi-speaker dialogue?

As of 1 October 2026: Google's Gemini TTS (2 speakers per request), ElevenLabs' Text to Dialogue (up to 10 voices per request), Microsoft's Azure multi-talker voices and open models such as Dia voice a dialogue in one pass. OpenAI and Amazon Polly take one voice per request, and studio apps such as Murf, WellSaid, Typecast, Narakeet and Kukarella's DialoguesAI voice it line by line.

Can I create multiple voices from one script?

Yes. In a dialogue model you mark each turn with its speaker. In a studio app you assign each line or block to a voice. In DialoguesAI the assistant writes the script with the speakers already assigned, or you create the speakers and assign the lines yourself.

Can AI assign different voices to different characters automatically?

Yes, but not by accent. Kukarella's assistant gives each speaker a voice by gender and age; you choose the accent. Most other tools ask you to pick every voice yourself.

Can Murf Studio be used for multi-speaker dialogues in storytelling?

Yes. Murf's help centre tells you to put each character's lines in separate blocks and pick a voice for each block (Murf). The voices take turns; they do not overlap.

How do I create a dialogue in ElevenLabs?

With Text to Dialogue, in the app's Dialogue mode or through the API: a list of turns, each with its voice — up to 10 voices per API request — and ideally no more than about 2,000 characters per request (ElevenLabs API reference).

Can two AI voices talk at the same time?

Only in a dialogue model, and only partly: ElevenLabs marks interruptions with punctuation, and Gemini can slip a short reaction into a speaker's turn. Tools that voice line by line, including Kukarella, play one turn after another.

How do I make a dialogue with cloned voices?

Clone each voice with its owner's permission, then cast the clones as speakers. In Google's Gemini API, cloned voices are generated one turn at a time (Gemini API docs). In Kukarella, a voice you cloned can play any speaker.

Can I make the dialogue in another language, or in two languages?

Yes. Name the language in your request, and give each speaker a voice in that language. A scene in two languages may not come back the way you asked; if it does not, write the lines yourself with Create Project. Give a speaker who switches languages a multilingual voice.

Is there an API for multi-voice dialogue?

Google's Gemini API, ElevenLabs and Microsoft Azure all have one. Kukarella has no public API; DialoguesAI is an app.

Can I use it in a paid course?

On Kukarella, commercial use comes with a paid plan (Pricing). On other tools, check the plan: ElevenLabs, for example, gives no commercial rights on its free plan.

How we checked

Kukarella's own facts — what the assistant does, how voices are cast, what the effects do, what each step costs — were read from Kukarella's own code as of 1 October 2026. Every other tool's facts come from that company's own documentation or pricing page, read the same day, unless we name another source; where two of a company's pages disagreed, we print the stricter one (ElevenLabs' 10 voices per request). Studies were read at the source: the paper, its official abstract or a public copy, and the numbers are the papers' own. The dialogue in the table and the requests in the planner are ours, written for this article.

About the author. Nazim Ragimov has been building Kukarella since 2019, first for himself and his team, then for the users whose feedback shaped almost everything in it. Most of them teach, and he wants Kukarella to delight them and their students: a lesson that makes the class say "Now we get it." He started these articles himself, with help from his wife, Anna; today he sometimes uses AI for research and first drafts.

Sources

All pages read on 1 October 2026 unless stated. Prices and limits change often; these are the figures of that day.

Studies

• Chi, Kang & Yaghmourian. Why students learn more from dialogue- than monologue-videos: Analyses of peer interactions. Journal of the Learning Sciences 26(1), 2017. PDF
• Chi, Roy & Hausmann. Observing tutorial dialogues collaboratively. Cognitive Science 32(2), 2008. PDF
• Muller, Bewes, Sharma & Reimann. Saying the wrong thing: improving learning with multimedia by including misconceptions. Journal of Computer Assisted Learning 24(2), 2008. DOI
• Bradlow & Bent. Perceptual adaptation to non-native speech. Cognition 106(2), 2008. PMC
• Baese-Berk, Bradlow & Wright. Accent-independent adaptation to foreign accented speech. Journal of the Acoustical Society of America 133(3), EL174–EL180, 2013. PMC
• Brekelmans, Lavan, Saito, Clayards & Wonnacott. Journal of Memory and Language 126, 2022. Oxford repository
• Zhang, Cheng & Zhang. The role of talker variability in nonnative phonetic learning: a systematic review and meta-analysis. JSLHR 64(12), 2021. PDF
• Ockey & French. From one to multiple accents on a test of L2 listening comprehension. Applied Linguistics 37(5), 2016. ERIC
• Kang, Thomson & Moran. TESOL Quarterly 53(1), 2019 (online 2018). DOI
• Chan, Lyons, Kon, Stine, Manley & Crossley. Learning and Instruction 67, 2020. DOI
• Craig & Schroeder. Journal of Educational Computing Research 57(6), 2019. ERIC
• Dinçer. Journal of Pedagogical Research 6(3), 2022. PDF
• Jing, Wu, Pi, Zhou, Zhang & Liu. The Journal of Experimental Education, online 28 December 2024. DOI
• Runge, Attali, LaFlair, Park & Church. Frontiers in Artificial Intelligence, 4 November 2024. Article
• Stivers and colleagues. Universals and cultural variation in turn-taking in conversation. PNAS 106(26), 2009. PMC
• Yuan, Liberman & Cieri. Towards an integrated understanding of speaking rate in conversation. Interspeech 2006. PDF
• Stolcke and colleagues. Dialogue act modeling for automatic tagging and recognition of conversational speech. Computational Linguistics 26(3), 2000. arXiv
• Bortfeld, Leon, Bloom, Schober & Brennan. Disfluency rates in conversation. Language and Speech 44(2), 2001. PDF
• Gilmore. A comparison of textbook and authentic interactions. ELT Journal 58(4), 2004. ERIC
• Wagner & Toth. Foreign Language Annals 47(3), 2014. ERIC
• Blau. More on comprehensible input: the effect of pauses and hesitation markers on listening comprehension, 1991. ERIC PDF
• Kirkland, Lameris, Székely & Gustafson. Interspeech 2022. ISCA
• Choi & Perrachione. Cognition 192, 2019. PDF
• Badcock & Chhabra. Frontiers in Human Neuroscience 7:114, 2013. Article

Exams, standards and teaching guidance

• IELTS Academic, Listening format. ielts.org
• Cambridge English: A2 Key handbook, B1 Preliminary handbook, B2 First handbook
• Pearson, PTE Academic Test Taker Handbook. PDF
• Council of Europe, CEFR Companion Volume, 2020. PDF
• W3C, Understanding WCAG 2.2 Success Criterion 1.2.1. w3.org
• Fiona Ledger, BBC, How to Write a Radio Play, undated, archived by Lancaster University. PDF
• Ross Thorburn, on using AI to create custom listening materials, Hong Kong TESOL blog, updated 19 June 2026. Blog
• Nik Peachey, on generative AI in the language classroom, British Council TeachingEnglish, undated. Article

The tools, from their own pages

• Google: Gemini API speech generation (updated 30 September 2026); Gemini 3.8 Flash TTS announcement; Cloud Text-to-Speech, Gemini-TTS; Audio Overviews launch, 11 September 2024; NotebookLM is now Gemini Notebook, 16 July 2026; Audio Overviews help
• ElevenLabs: Text to Dialogue; API reference; Prompting best practices; Professional Voice Cloning
• Microsoft: SSML voice and sound; SSML structure; Speech service release notes; Disclosure for voice talent; VibeVoice on GitHub
• OpenAI: Text to speech guide
• Murf: Project with multiple characters · Narakeet: Multiple voices · Typecast pricing · WellSaid pricing
• Nari Labs: Dia, Dia2
• Artificial Analysis, text-to-speech leaderboard

Stories and reporting

• TechCrunch, Meta acquires voice startup Play AI, 13 July 2025. Article
• Simon Willison, notes on NotebookLM's Audio Overviews and the Hard Fork interview, 29 September 2024. Blog
• Sesame, Crossing the uncanny valley of conversational voice, 27 February 2025. Research post
• Fortune, the Arup deepfake fraud, 17 May 2024. Article
• CNN, Finance worker pays out $25 million after video call with deepfake 'chief financial officer', 4 February 2024. Article

Kukarella

• DialoguesAI and Pricing, 1 October 2026. How the assistant casts voices, what the effects do and what each step costs were read from Kukarella's production code the same day.