To narrate an e-learning course with an AI voice, write the script so that each slide or screen is one short paragraph, choose one voice for the whole course, voice it, and download one audio file per paragraph. Then put each file on its slide in Storyline, or in its block in Rise, and add the script as captions. When a policy changes next year, you edit that one paragraph, voice it again and swap one file. A 20-minute module is about 18,000 characters of script; in Kukarella that is about 18,000 credits, a small part of a US$21 month.
But the voice is the easy part. What decides whether the course works is what happens around it: words the voice says wrong, captions that drift from the audio after an edit, and a voice that can disappear from the tool you made it in. I build Kukarella, one of the tools in this guide, and it has gaps of its own: it gives you audio files, not a plug-in for your authoring tool, and no caption file. I say where each option is weaker as we go, including when you should hire a person.
The whole path in seven steps:
- Decide where the voice comes from: your authoring tool's own AI voice, a separate voice tool, or a human narrator.
- Write the script for the ear, one paragraph per slide, with acronyms and numbers spelled the way they should sound.
- Choose one voice and one accent for the whole course, and write its name down.
- Voice it and listen to every paragraph before you build.
- Put the audio in Rise, Storyline or Captivate, and add captions or a transcript.
- Keep a narration log: which paragraph is on which slide, in which file, in which voice.
- Update it when the content changes: one paragraph, one file, one republish.
Should you use your authoring tool's AI voice, a separate voice tool, or a person?
Short answer: if your whole course lives in Articulate 360 with its AI package, start with its built-in voice; it costs nothing extra there. Use a separate tool when you need a voice and an accent you control across several tools, role-plays with several speakers, or files you can keep and reuse. Hire a person for the messages where learners must trust a human. But every built-in voice ties your narration to one tool and one subscription.
Something surprised me when I read the documentation of the big authoring tools: their AI voices come from the same provider. Articulate names ElevenLabs as its AI voice subprocessor (Articulate); Adobe asks closed networks to allow api.elevenlabs.io for Captivate's AI voices (Adobe); iSpring says its AI voices are "powered by ElevenLabs" (iSpring); and so does TechSmith for Camtasia (TechSmith). So the real choice is not which AI. It is whether the narration lives inside your authoring tool or outside it.
Your authoring tool's own AI voice
- Best when: the course is built and kept in one tool, by people who all have its AI plan.
- What it costs: Articulate's text-to-speech comes with the Articulate 360 AI package, US$1,749 a seat a year for teams (Articulate); its credit table lists no charge for it (Articulate). iSpring Suite AI, US$1,290 an author a year (iSpring), includes 200,000 ElevenLabs characters every 30 days, shared with iSpring Cloud AI (iSpring).
- What it does well: the audio lands on the slide or in the block, with a transcript (Rise) or captions (Storyline), and on the AI plans Articulate Localization translates AI narration along with the course, using credits (Articulate).
- The catch: if your AI access ends, Rise keeps the audio but it "becomes a regular audio file" you can no longer edit with the AI (Articulate). A Rise script is capped at 5,000 characters, depending on the model (same page). And the voice you chose can be withdrawn: 19 voices leave Articulate on 1 January 2027, and every clip made with them must be generated again with a replacement that is similar, but whose "tone and style may differ" (Articulate).
A separate voice tool (Kukarella, ElevenLabs, Murf, WellSaid and others)
- Best when: you work in more than one tool or with an LMS that only takes files; you need an accent your authoring tool's library does not have; you make role-play dialogues with several voices; or your team includes people without the authoring tool's AI plan.
- What it costs: at Kukarella most voices use 1 credit a character (Kukarella), and the Prime plan is US$21 a month for 300,000 credits (pricing). ElevenLabs' Creator plan is US$22 a month (ElevenLabs).
- What it does well: the audio is yours as ordinary files; one voice can serve every course, in every tool.
- The catch: you place the files by hand, and Articulate Localization does not translate audio you import: imported audio is "not translated automatically but can be customized per language" (Articulate). Kukarella makes no caption file, has no SCORM export and no Articulate plug-in.
A human narrator
- Best when: a leader speaks to staff, the topic is sensitive (harassment, grief, safety after an accident), or the voice is part of the brand.
- What it costs: a UK agency gives a guide price of £600 plus VAT for up to 1,500 words of e-learning (voiceovers.co.uk); a US rate page based on the GVAA guide gives US$30–55 per finished minute for e-learning (VoiceCrafters).
- What it does well: warmth, judgement and trust that learners still rate higher (next section).
- The catch: every later change is a booking. One UK price list charges £150 for a pickup of up to 150 words (Outspoken Voices), and if the narrator is not available, the new line will not match the old ones.
This quick check puts three real situations to you, and says which way I would go in each and why.
Does an AI voice hurt learning?
Short answer: with a good modern voice, the research mostly finds that people learn as much as from a human voice. But learners still rate a human voice higher, and the old, flat synthetic voices did worse in the early studies. So use AI for procedures and content that changes, and a person where trust is the point.
The idea that a machine voice teaches worse comes from a well-known 2003 study. Richard Mayer and colleagues found that learners solved new problems better after a human voice than after a machine voice, and after a standard accent than after a foreign one (Mayer, Sobko & Mautone, 2003). That machine voice was a synthesizer of the early 2000s.
When the test was run again with modern voices, the gap closed. Scotty Craig and Noah Schroeder found that learners who heard a modern text-to-speech engine "in most respects" were "not statistically different" from those who heard a human, in what they learned and how they judged the voice (Craig & Schroeder, 2019). Two later studies found the same learning but a difference in feeling: non-native listeners rated a human voice higher on most scales (Davis, Vincent & Park, 2019), and the quality of a voice changed how much learners trusted the speaker (Chiou, Schroeder & Craig, 2020). Accents show the same pattern: a foreign-accented instructor did not change what students learned, but they rated the instructor as less effective (Sanchez & Khan, 2016).
The newest data comes from a vendor, so read it as such. In a study TechSmith published in October 2026 (it sells AI voices), 768 full-time workers watched one of four versions of a training video: those who heard a high-quality AI voice answered 71% of quiz questions correctly, against 61% for a high-quality human voice (TechSmith, 2026). The human version scored lowest of the four, which is a reason to read it with care. Its own advice matches mine: AI for procedural and often-updated content, people for leadership and sensitive topics.
Tell learners. Florida Gulf Coast University suggests one plain line: "This audio was created from instructor-written text using an AI text-to-speech tool" (FGCU, 2026). I would put a line like it on the course's first screen.
How do you write a script that an AI voice reads well?
Short answer: write for the ear: short sentences, one idea per paragraph, one paragraph per slide, and every acronym and number spelled the way it should sound. But a spelling that fixes the sound can break the caption, so keep the written words and the spoken words apart where they differ.
One paragraph per slide is the rule that pays for itself later. In Kukarella the paragraph is the unit of everything: each paragraph is voiced, stored, downloaded and charged on its own. When you change a sentence, only that paragraph is voiced and charged again; the others keep their audio at no cost. A paragraph that covers one slide becomes one file for that slide, and a fix touches one file. A paragraph that covers three slides means you cut audio by hand, every time.
Spell for the voice. A voice reads what it sees. In Kukarella there is no pronunciation dictionary; you have two tools. One is Say as in the Effects panel, which tells a voice to spell out letters or read a number as a cardinal, an ordinal or digit by digit, on the voices that support it. The other is the spelling itself: write an acronym with full stops between the letters (C.R.M.) so the voice spells it, and write a phone number or an extension digit by digit.
Then keep the captions honest. If the narration says "C. R. M." or "one eight hundred", the caption should still read "CRM" and "1-800". Keep the script as two columns for the few lines where they differ: what the learner reads, and what the voice reads.
Write it to be heard:
- One idea per sentence, under about 20 words. A listener cannot scroll back.
- Say the action first: "Click Submit, then check the email that arrives", not "Once the form, which you completed in step 2, has been submitted…".
- Leave a pause where you want a learner to look. Kukarella's pauses run from half a second to 90 seconds, and the pause is saved inside that paragraph's file, so it travels with the slide.
- Read numbers aloud as you write them: "about a third" is easier to hear than "33.4%".
- Time it. At 150 words a minute a 20-minute module is about 3,000 words. Voice one paragraph and time it with your voice before you write the rest.
If your script starts as a Word file or a PDF, Kukarella's Media button takes DOCX, PDF and TXT and puts each line in a paragraph of its own, up to 500 paragraphs a project. It does not take PowerPoint files: copy the speaker notes into a document first.
Which voice and accent should a course use?
Short answer: one voice for the whole course, in the accent of the people who take it. A French course for staff in Quebec wants a French (Canada) voice; an English course for a team in Sydney might want an Australian one. But a voice can be retired by its provider, so write down which voice you used and test a replacement before you need it.
One voice, many modules. Learners hear a course as one thing. In Kukarella you choose the voice per project, so with one project per module you pick the same voice each time; nothing links them, which is why the narration log in this guide has a column for the voice.
A native voice for a real accent. Kukarella's Voice Library lists each language by country, and shows two kinds of voice: native voices, which belong to one language and carry that country's accent, and multilingual voices, which speak many languages with one base accent ("They speak French too, but may not have this accent," the library says). For a regional accent, pick a native voice of that country. Kukarella has 2,000+ voices; you need one.
A voice can leave. Voices are made by providers, and providers retire them; Articulate's January 2027 list is the latest example. When one of ours is retired, you can send us a few recordings you made with it, and we can make a clone that sounds very close to it; but I cannot promise that any voice, ours or anyone's, will be there forever. A note of the voice's name and settings is what lets you match it or replace it cleanly.
For several voices in one scene — an agent and a customer, a nurse and a patient — use a dialogue tool rather than one narrator. In Kukarella that is DialoguesAI, where each speaker has a voice of their own; AI Role-Play Dialogues for Staff Training covers it. A 2025 study found lower cognitive load and better retention when learners heard three AI voices, one per subtopic, rather than one (Liew et al., 2025).
How do you make the narration in Kukarella, step by step?
Short answer: paste or import the script into a Text to Voice project, choose the voice, press Play, listen, and download By paragraph to get one MP3 or WAV per paragraph in a ZIP. But listen to every paragraph before you download: the voice will read exactly what you wrote, including the mistakes.
- Open Text to Voice and start a project for the module. Name it after the module, for example M03 Data protection: the name goes on the ZIP you download, so each module's files stay apart.
- Bring the script in. Paste it, or use Media to import a DOCX, PDF or TXT file. Each line becomes a paragraph. Make sure each slide is one paragraph.
- Choose the voice in the Voice Library: the language, the country, then a native voice. Play a sample paragraph in two or three voices before you decide.
- Fix the reading. Select a word and use the Effects panel: a pause, the speed, or Say as for letters and numbers. Not every voice takes every effect; the panel shows what the chosen voice supports.
- Press Play and listen to the whole module. Each paragraph is voiced once and stored. A paragraph whose text, voice and settings you do not change is not charged again.
- Download. Choose MP3 or WAV, then By paragraph. You get a ZIP with one file per paragraph, numbered in order from 0 and named with the paragraph's first 20 characters and the voice, such as 03. Click Submit to send-Aria.mp3 for a paragraph that begins Click Submit to send the form. File 00 (or 0) is your first slide. Every paragraph also has its own Download button, which you use after a fix; its file is named after the project and its version (M03 Data protection--v7.mp3), so rename it to the name in your log. And if you add or remove a paragraph, the files after it get new numbers in the next ZIP: rename from your log rather than trusting the numbers.
- Download the text too (TXT or DOC) and keep it with the audio. It is your caption text and your transcript.
This is the Download window of a six-paragraph module, with By paragraph chosen:
Kukarella's Download window: MP3 or WAV, One file or By paragraph, and the text as TXT or DOC.What you will not get: a caption file, loudness matching across voices, or a direct export to an LMS. The files are plain audio. If you mix voices from different providers, check the levels side by side.
How do you put the audio into Rise, Storyline or Captivate?
Short answer: in Storyline, one file goes on each slide's timeline; in Rise, one file goes in each audio block; in Captivate, one file per slide. Then add captions or a transcript from your script. But no voice tool outside your authoring tool can place the files for you.
Storyline 360. Insert the paragraph's file on its slide; it lands on the timeline, where you sync animations to it. Storyline imports MP3, WAV, M4A and several other formats (Articulate). For captions, Storyline imports SRT and VTT files, and picks up a caption file automatically when it has the same name as the audio and sits in the same folder (Articulate). Kukarella makes no caption file, so you write one from the paragraph's text, or use Storyline's caption editor.
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.
Rise 360. Rise places audio by block: an audio block, or the audio inside some interactions. You can import a file of up to 5 GB (Articulate). Put the paragraph's text right under its audio block as the transcript.
Captivate, iSpring and an LMS. Captivate and iSpring take ordinary audio files per slide. An LMS takes the published course, so the audio always goes in through the authoring tool first.
Captions and transcripts are not optional. Under WCAG, audio heard on its own needs a text alternative, and audio timed to slides or animation needs captions (W3C, 1.2.1; W3C, 1.2.2). In the United States, public bodies, public universities included, must meet WCAG 2.1 AA from 26 April 2027, or 26 April 2028 for smaller ones, after the Department of Justice moved the dates in April 2026 (University of Arizona). And check the captions after every edit: even Articulate had a bug, fixed in September 2026, where captions for AI narration did not update after the narration was edited (Articulate).
How do you change one line after the course is live?
Short answer: change the paragraph, voice it again, download that one file, replace it on its slide or block, fix the caption, and republish. With a narration log it takes minutes. But changing the voice, or a setting that applies to the whole project, voices every paragraph again and charges for all of them.
This is where AI narration earns its place. Training goes out of date: a form changes, a policy is revised, a product gets a new button. With a human narrator, a two-sentence change means a pickup session, a fee, a few days, and a hope that the voice matches. In a case study published by WellSaid, the voice tool they moved to, the manager of Microsoft's internal enablement team, Doug Shaddick, puts it plainly: narration was the slowest step, and "that can take two or three hours for a single snippet" (WellSaid). He also says the honest part: "If WellSaid went away tomorrow, we would struggle." Whatever tool you use, keep the script as the master copy.
The update, step by step in Kukarella:
- Find the paragraph in your narration log: module, slide, file name.
- Edit the text in the module's project. Only that paragraph will be charged.
- Press Play on it and listen.
- Download that paragraph with its own Download button, and rename the file to the name your log uses.
- Replace the audio on its slide in Storyline, or in its block in Rise.
- Update the caption and the transcript from the new text.
- Republish to your LMS, and note the date in the log.
What a change costs. A paragraph of about 80 words is about 480 characters; at 1 credit a character, five changed paragraphs are about 2,400 credits, about 17 US cents' worth of a Prime month. The same change from a narrator on the UK price list above is a £150 pickup at least, plus VAT.
What to avoid: do not change the voice, the speed of the whole voice or the pauses for the whole project to fix one line. Those settings go into every paragraph, so every paragraph is voiced and charged again. And keep the old project: each conversion saves a version you can restore.
Paste a module's script below to make its narration log: one row per paragraph, with the file name Kukarella will give it, its length and its credits. Paste the changed script under it, and it tells you which files to replace and what the change costs.
What does narrating a course cost, the first time and at every update?
Short answer: in Kukarella a 20-minute module in most voices is about 18,000 credits, so Prime's 300,000 credits a month cover about 16 such modules, and a fix costs only the paragraphs you change. A human narrator costs from several hundred to over a thousand dollars or pounds for the module, and a minimum fee for every change. But the AI figure is the voice alone: your time to write, listen, place files and check captions is the same either way.
| First narration of a 20-minute module | A change of two sentences | |
| Kukarella, most voices | about 18,000 credits (US$21 a month buys 300,000) | the changed paragraphs only, about 1,000 credits |
| Built into Articulate 360 AI | no listed charge on the AI plan (US$1,749 a seat a year) | no listed charge; regenerate the clip |
| Human narrator (a UK agency's guide price) | about £1,200 plus VAT for 3,000 words | from £150 for a pickup, and a booking |
Assumptions: 150 words a minute and about 6 characters a word, so a 20-minute module is about 3,000 words or 18,000 characters; two sentences sit in two paragraphs of about 80 words. The UK figure is voiceovers.co.uk's guide price of £600 for the first 1,500 words plus £100 for each further 250 words, before VAT; the pickup is Outspoken Voices' price for up to 150 words. Prices read on 19 November 2026.
Three things the table hides. First, some voices cost more: in Kukarella some voices use more credits a character, and the editor shows the count before you press Play. Second, commercial use: Kukarella's paid plans include it, and audio made on a paid plan keeps its rights after you stop paying; the free plan does not include commercial use. Third, the people: in Kukarella you can share a project with up to 10 colleagues as editors or viewers, one editor at a time; viewing is free, and each editor needs a plan or a seat you add to yours.
If you are weighing a human narrator, or cloning your own voice, for a whole course, Voice Cloning vs Voice Actors: What a Course Costs has a calculator for it.
Can you make the same course in another language?
Short answer: yes, but treat each language as its own course. Translate the script, have a native speaker check it, and choose a native voice of the country where it will be heard. But do not expect a one-click translation and dubbing: in Kukarella there is none.
If your course is in Articulate and you use its AI narration, Articulate Localization translates the narration with the course, which is the strongest reason to stay inside the tool for a multilingual course (Articulate). Outside it, the work is: translate the script (in Kukarella, Ask AI inside the project can translate the text; it uses credits), check it with someone who speaks the language, keep one paragraph per slide in every language so the files still match the slides, and choose the voice for the accent. Which accent learners need, and where to find it, is in Text to Speech Accents: Which One Learners Need.
A note for Quebec: the Charter of the French Language requires written communications with employees, written training materials included, to be in French, according to a 2023 summary by the law firm Osler (Osler); check how it applies to your organisation. It speaks of written material; for narration, a French (Canada) voice is simply what your learners will expect.
Can a narrated section become a short video?
Short answer: yes. In Kukarella, a Text to Voice project can become an MP4 with a picture for each moment and captions, through Create Visuals → Video (Make video): for one minute of voice, about 38 visuals, and the window estimates about seven minutes of making. But it is made for a one- to three-minute recap or trailer, not a whole module: it stops at 10 minutes, nobody's lips move, and each picture must be checked before learners see it.
A short video is useful at the edges of a course: a 60-second introduction that sells the module, a recap at the end, or a scenario learners watch before they practise. Make video takes the voice you already made, draws a picture for each part of it, adds captions you can switch off, and gives you an MP4 to place in Rise or Storyline like any video. A free account can make one video, of up to 0:45 of voice, once, with a Made with Kukarella.com mark on every frame; more videos come with a paid plan and cost visuals. It opens in the vertical 9:16 shape, so choose Horizontal 16:9 for a course.
This is a video I made from a dialogue with Make video:
What Make video is, who it is for and how it compares with other tools is in Turn a Dialogue Into a Video: Make Video Explained. Every screen, step by step, is in How to Turn a Script Into a Video for Class.
Frequently asked questions
Can I use AI narration in a course I sell? With a paid plan, yes: Kukarella's paid plans include commercial use, including courses you sell and work for clients. The free plan does not include commercial use. ElevenLabs draws the same line between its free and paid plans (ElevenLabs).
Does Kukarella connect to Articulate, SCORM or my LMS? No. It gives you MP3 or WAV files, whole or one per paragraph, which you place in your authoring tool. There is no plug-in and no public API.
Can the AI write the narration for me? Kukarella's assistant can draft a script from a topic, a level and a length, and pick a voice; you then press Play to voice it. Check every fact it writes: it does not know your policy.
Should I clone the trainer's own voice? It keeps a familiar voice in every module and every update, but it needs the trainer's consent, and the question of what happens when the trainer leaves. Georgia Tech's College of Lifetime Learning has the person whose voice is cloned sign a media release and fixes outdated lines by regenerating them; its team says the audio pipeline cut production and editing time "by over 60 per cent", a figure they give without a method (Times Higher Education Campus, 2026).
What if a voice I used is retired? Find a close replacement, voice the whole module again in it, and replace every file; a course in two voices sounds like two courses. This is why the log keeps the voice's name.
