Facts, menus and prices checked in November 2026.
To make the same lesson in another language, translate the script, not the recording; have someone who speaks that language check the translation; then voice it again with a voice of the country where it will be heard, and re-time it, because the new version will not be the same length. Expect a narration to run roughly 10% longer in French or Spanish and about 20% longer in German than in English, and the written script, in European languages, to grow by about 30%. Treat each language as its own lesson, made from the same source.
But the step that goes wrong is rarely the voice. It is the translation: machine translation is very good in Spanish and much weaker in smaller languages, and nobody on your team can hear a mistake in a language they do not speak. In Kukarella, where we make the audio, there is also no single button that translates and re-voices a lesson. You translate with Ask AI, one paragraph at a time, or the whole lesson at once if it is short, then choose the new voices yourself. Each section below says what holds for any tool, then what to do in Kukarella.
In this guide:
- Is it worth making a second language?
- Five ways to do it, and what each costs.
- Step 1: freeze the lesson and write a glossary.
- Step 2: translate the script, then have it checked.
- Step 3: choose the voice for the new language.
- Step 4: re-time it: the new version is longer.
- How to do it in Kukarella.
- One lesson that mixes two languages.
- Captions and video in the new language.
- When the lesson changes: keep the versions in step.
- Frequently asked questions.
Is it worth making a second language?
Short answer: usually yes. People understand more in their own language, even when they speak the other one well, and in the one controlled test we found, the same AI-made video dubbed into the learners' own language did better than the English original. But only if the learners want it: some prefer to study in the language of their field.
The scale is large. UNESCO's education report said in 2016 that, by one estimate, as much as 40% of the world's population has no access to education in a language they speak or understand, and the World Bank put the share of pupils in low- and middle-income countries who must learn in another language at 37% (2021). Those are school figures, but the same logic runs through workplace safety: in the United States, OSHA requires employers to train workers "using both a language and vocabulary that the employees can understand" (OSHA, 2010).
The closest evidence for a narrated lesson comes from a 2026 study at the University of Valencia. Forty native Spanish speakers with intermediate to advanced English (B1 to C2) watched two-minute videos in which two AI presenters gave facts about a place. They watched some in the English original and some dubbed into Spanish, with voices made by the same AI video tool. They answered more comprehension questions correctly after the Spanish version, 8.93 out of 10 against 8.15 (Romero-Ortells and colleagues, 2026). The authors note that learners with weaker English were not tested.
The best story I know of a whole course in many languages is the World Health Organization's. When COVID-19 began, its OpenWHO platform built an introductory course in about three weeks; by 25 March 2020 it was in 13 languages, and more people had enrolled in Spanish (51%) than in English (38%) (Utunen and colleagues, 2020).
The "but" comes from a study of South African students who heard a lecture in English with subtitles in their home language, Sesotho: they often skipped the Sesotho subtitles, and they scored no better than students who had English subtitles or none (Kruger, Hefer and Matthew, 2014). Before you translate, ask your learners which language they study in.
Five ways to do it, and what each costs
Short answer: translating the script and voicing it with text to speech is the cheapest way you control; a human translator and voice actor is the best and the dearest; AI dubbing keeps your own voice but starts from the audio. But every route that makes audio needs someone who speaks the language to check it.
Prices below were read on each company's page in November 2026; they change often.
1 · A human translator and a voice actor
- What it costs: one UK agency's published card gives 13p–18p a word to translate, 40% more for a second translator's proofread, and £275–£500 an hour for a voice artist, plus studio and editing (Matinée Multilingual). For a 10-minute narration of about 1,500 words, that comes to roughly £690–£1,020 per language, by my arithmetic.
- Best for: a flagship course, or a voice that has to act.
- But: it is the slowest route, and every later change to the script means booking the actor again.
2 · AI dubbing of the video, in your own voice
- What it costs: ElevenLabs' dubbing works out at about $0.50–$0.60 for each minute of video, per language, on its paid plans; Rask AI at about $1.00–$1.60 a minute on monthly plans, less if paid yearly; HeyGen from $29 a month, with lip-sync. YouTube dubs eligible videos automatically, and its help page names no charge.
- Best for: a video where your face and voice are the lesson.
- But: it starts from your recording, so a mistake in the transcript or the translation travels into the dub unless someone who speaks the language reads it first; YouTube's automatic dubs cannot be edited at all; and cloning a voice needs the speaker's consent.
3 · Your course tool's own translation
- What it costs: Articulate 360 seats are $1,449–$1,749 a year, and its Localization charges 2 credits per source character per language, about $0.01 a character on its published credit price.
- Best for: a course already built in Articulate with its AI narration, which it re-voices in the new language.
- But: recorded narration and imported audio are "not translated automatically" (Articulate), so audio made elsewhere has to be made again.
4 · Subtitles only
- What it costs: nothing, on YouTube and most video tools.
- Best for: learners who already follow the spoken language.
- But: they read while they watch, and YouTube itself warns that automatic captions may misrepresent what was said.
5 · Translate the script, then voice it with text to speech
- What it costs: the least; in Kukarella, credits for the translation and for the voice (the Kukarella section has the rates).
- Best for: a narrated lesson, a slide deck, a listening dialogue, a lesson you will change again.
- But: you re-time the audio to your slides yourself, nobody lip-syncs it, and the check by a speaker of the language is your job. This is what Kukarella does, and the rest of this guide follows this route.
Step 1: freeze the lesson and write a glossary
Short answer: translate a finished lesson, not a draft, and decide a few words before anyone translates a sentence. But a glossary only helps if the translator and the voice both follow it.
Every change you make to the source after translation starts has to be made again in every language, so finish the original first: the script, the on-screen text, the names. Then write a short glossary, one line per decision:
- Terms your field uses, with the one translation you want, and the ones you do not ("lockout", "near miss", a product name you never translate).
- How the lesson addresses the learner. French, German, Spanish and many other languages have a formal and an informal "you" (vous or tu, Sie or du, usted or tú). Choose one and keep it in every line, spoken and on screen.
- Which country's version. Spanish for Mexico or for Spain, French for Canada or for France, Portuguese for Brazil or for Portugal: the words differ, and so do the voices.
- Numbers, dates and units: "1,500" is "1.500" in German; a date may need its day first; miles may need to be kilometres.
Step 2: translate the script, then have it checked
Short answer: machine translation, or an AI assistant, gives a good first draft in big languages; a person who speaks the language then reads every line. But you cannot skip that read for a language you do not speak, because you will not hear what is wrong.
The best-known test is medical. In 2019, doctors at the University of California, San Francisco ran 647 sentences of emergency discharge instructions through Google Translate. 92% of the Spanish sentences came back accurate and 81% of the Chinese ones; one Chinese sentence turned "hold the kidney medicine" into "keep taking kidney medicine", the opposite instruction (Khoong and colleagues, 2019). A 2021 study of seven languages found the gap wider still: the meaning survived in 94% of Spanish translations and 55% of Armenian ones, where ibuprofen became an "anti-tank missile" (Taira and colleagues, 2021). Newer AI tools have narrowed the gap in big languages but not closed it: in a 2024 test of children's discharge instructions, ChatGPT-4 and Google Translate matched professional translators in Spanish and Portuguese, but a third of ChatGPT's Haitian Creole translations had errors that could harm a patient (Brewster and colleagues, 2024).
So the order is: translate, then have it read. The reader can be a colleague, a learner's parent, a former student; the international standard for translation services, ISO 17100, makes a second person's revision part of the job. Ask them to read the translation against the original, aloud, the way the learners will hear it.
My own limit is the same as yours. When I listen to voices in other languages, I can judge English and Russian, the two I know. In every other language I cannot hear a wrong stress, and I have to ask someone who can.
Step 3: choose the voice for the new language
Short answer: for a language version, choose a native voice of the country where the lesson will be heard. Choose a voice that speaks many languages only when the same presenter across every version matters more than the accent. But hear it first, for at least half a minute.
A native voice speaks one language with the accent of one country: Mexican Spanish, Canadian French. A multilingual voice reads many languages, which is what you want when one voice must stay the same in every version, or when one script switches language. But it usually keeps one base accent: in my own listening it sounds fine at first, and after about fifteen seconds you hear a slip in stress or pronunciation. Which accent your learners need, and how to find it, is in Text to Speech Accents: Which One Learners Need.
Listen below: the same paragraph of a lesson in English, then in Russian read by a native Russian voice, then in Russian read by a voice that speaks many languages, an American English voice. The Russian was translated with Ask AI, as in the Kukarella section further down. Listen to the stress and the vowels of the third, then decide whether that difference matters to your learners. I chose Russian because it is the one language besides English I can judge myself. The English runs 22 seconds; the native Russian, 28.
First, the English original:
The lesson paragraph in English: Ryan, a native English (UK) voice
Then the Russian, read by a native voice of Russia:
The same paragraph in Russian: Dmitry, a native Russian voice
And the same Russian, read by a voice that speaks many languages:
The same Russian paragraph: an American English voice that speaks many languages
Your own voice in every language. Many teachers ask whether they can clone their voice and have it speak the other languages. In Kukarella a quick clone speaks 16 languages and a professional clone 40 (Clone a Voice), so learners hear the same person in every version. But expect your own accent to come with you: a clone is a voice that speaks many languages, and you will be the one listener who cannot judge how you sound in a language you do not know. Ask a native speaker, as in Step 2.
Step 4: re-time it: the new version is longer
Short answer: plan for about 10% more time in French and Spanish and about 20% more in German, and about 30% more text on screen in European languages. But these are averages from a small study of human readers; time each paragraph by ear in the new language.
Two kinds of length change. Written text grows: the W3C's guide to text size, based on IBM's design guidelines, says text over 70 characters runs about 30% longer when English is translated into European languages, and short labels can double. Speech changes pace: in a study of seven languages read aloud by native speakers, Spanish speakers said 7.82 syllables a second and English speakers 6.19, but each Spanish syllable carried less, so the same content took about 10% longer to say (Pellegrino, Coupé and Marsico, 2011). From their figures, the same content takes roughly 9% longer to say in French, 10% in Spanish, 20% in German and 15% in Mandarin than in English, by my arithmetic.
What that means for a lesson: a slide timed to 40 seconds in English will need about 44 in Spanish; a caption that fitted on two lines may need three; a dialogue cut to a two-minute listening task may run to two minutes twelve in Spanish. Keep one paragraph per slide in every language, so each audio file still matches its slide (How to Narrate Your Slides With an AI Voice shows why).
The planner below does the arithmetic for your lesson: paste the script or give its length, choose the languages, and it shows how long each version will run, what it costs in Kukarella credits, and what to decide for each language before you translate.
How to do it in Kukarella
Short answer: start a new project for the new language, paste the lesson, translate it with Ask AI, then change the voice and listen. But there is no "translate and dub" in one click, a long lesson translated in one go comes back cut short without a warning, and in DialoguesAI replacing the whole text takes every line's speaker away.
Ask AI on one paragraph: Translate to Russian, then Replace source text.- Start a new project for the new language, in Text to Voice for a narration or DialoguesAI for a dialogue, and paste the original lesson. There is no "Duplicate" for a project, so this is how you keep the original as it is. In DialoguesAI, Import actors brings over the cast of the original project. But pasted lines carry no speaker: give each line its actor before it can play.
- Translate. For one paragraph, click Ask AI on it, type the request ("Translate to Spanish (Mexico), formal usted"), send, and choose Replace source text: the translation takes the paragraph's place, and in DialoguesAI the line keeps its speaker. But in DialoguesAI, Ask AI appears on a line only after that line has been voiced, and on a phone the "Translate to" suggestion is hidden (you can type the same words). For the whole lesson at once, Ask AI in the toolbar opens Ask KukaGenie with all of the project's text. Its answer stops at roughly 3,000 English words, fewer in many other languages, and it stops without a warning: for a longer lesson, translate paragraph by paragraph, or split the lesson into several projects. Replace original text puts the translation in as plain text, so pauses and effects go, and in DialoguesAI every line loses its speaker; Add text to project adds it below the original.
- Change the voice. Open the Voice Library, choose the language and country, and pick from Native voices first. In Text to Voice one voice reads the whole project, so select the new voice and unselect the old one. In DialoguesAI, change each actor's voice once and all their lines follow.
- Listen, and fix what sounds wrong. If a name or a term is said wrongly, spell it as it should sound: the effects panel's Say As covers letters, numbers, ordinals and fractions only, and there is no pronunciation dictionary.
- Download MP3 or WAV, as one file or By paragraph, one file per slide.
What it costs. A translation by Ask AI costs 1 credit for every 5 characters you send, the lesson text included; the translated answer is not charged. But each further message in the same chat sends the whole conversation again, the earlier answers included, and is charged for it, so start a new chat for each paragraph or each project. A free account has 10 Ask AI requests in each app and a guest 3. Voicing costs 1 credit per character for most voices, 2 for some. And on the free plan the audio has no commercial rights, which matters for a course you sell or use at work.
The Voice Library after choosing Russian: Native voices first, then Multilingual voices.If you ask the AI assistant that writes scripts for a new script in another language instead, with the old lesson attached, you get a new lesson built on the old one, not a line-by-line translation. That is useful for a listening dialogue at a new level (AI for Language Teachers: Listening Dialogues); for the same lesson in a new language, translate.
One lesson that mixes two languages
Short answer: for a script that switches language, put it in DialoguesAI and give each language its own line and its own native voice; use one voice that speaks many languages only if the switch happens inside a sentence. But expect that one voice to carry one accent in both languages.
Language teachers often need this: an English explanation, then the Spanish phrase, then English again. In DialoguesAI each line has its own speaker, so put the Spanish phrase on its own line and give it a speaker with a Mexican or Spanish voice; the learner hears each language spoken natively. Text to Voice cannot do this: one voice reads the whole project. When the switch happens in the middle of a sentence ("The word for train is el tren"), one voice that speaks both languages reads it without a break, and that is the job those voices are best at, as long as the accent of the second language does not matter to your lesson.
I recorded the video below in 2025: one script that moves through French, German, Russian, Japanese and English, voiced in Text to Voice by voices that speak many languages. The screens have changed since, but the idea is the same: listen to how one voice carries its accent from language to language, then decide whether your lesson can live with it.
Captions and video in the new language
Short answer: make captions from the new audio, not from the original. But a translated transcript keeps the original's timing and language in Kukarella's caption file, so translating the transcript is not enough.
In TranscribeHub, Ask AI changes the text in the editor, but the SRT caption file is built from the original transcription, in the original language. So transcribe the new language's audio file, check it, and download its SRT: How to Make an SRT or VTT File From a Video shows how.
A voiced project in the new language can also become a short video with Make video: the captions are the project's own text, timed to the voice, so they are already in the new language (Turn a Dialogue Into a Video: Make Video Explained). But check the captions in a language with its own script: fonts for Latin, Greek, Cyrillic, Vietnamese, Japanese, Korean, Hindi and Chinese are built in; Arabic, Hebrew and Thai are not, so check those captions before you share the video.
When the lesson changes: keep the versions in step
Short answer: change the original first, then the same paragraph in every language, and re-voice only that paragraph. But keep a list of which language versions exist, or one of them will be left behind.
This is where one paragraph per slide pays for itself again. When a rule or a figure changes, you edit one paragraph in the original, translate that paragraph, have it read, and re-voice it in each language project: a few minutes per language, not a new recording. With a human voice actor, the same change means a new booking.
Frequently asked questions
Can Kukarella translate and dub my video?
No. Kukarella does not dub a video. You can transcribe the video in TranscribeHub, translate the text with Ask AI, and voice it with a new voice; putting the new audio back on the video is a job for your video editor.
Does translating cost credits?
Yes: 1 credit for every 5 characters you send to Ask AI, the lesson text included. The translated answer is not charged, but a follow-up in the same chat sends the earlier answers again and is charged for them.
How many languages can Kukarella voice?
The Voice Library has 2,000+ voices in more than 150 languages and accents. Not every language has a native voice for every country; the Voice Library shows how many voices each accent has.
Can one voice read the lesson in every language?
Yes, a voice that speaks many languages, or your own cloned voice. But it usually keeps one accent in every language, so a native speaker should hear it before your learners do.
Will my cloned voice keep my accent in another language?
Expect it to. A quick clone speaks 16 languages and a professional clone 40, in your voice; how natural it sounds in each depends on the language, and you are the one person who cannot judge it.
Can I use the translated audio in a course I sell, or at work?
On a paid plan, yes, as long as you hold the rights to the lesson text: paid plans include commercial use. Audio made on the free plan has no commercial rights.
