Producing in another language

A course can be produced in a language other than English. This page is precise about what that does and does not mean, because “multilingual” is a word that hides a lot.

What the language applies to

The course output. Everything a learner sees or hears:

• The spoken narration.
• Each lesson’s opening line, which is narrated, or spoken on camera in the lesson an avatar presents where one is set.
• The text on the slides.
• The subtitles.
• The student workbook.
• The quiz.
• The listing fields.
• The course opening, where one has been produced in that language. Arabic has its own. A Malayalam course currently opens with the English one, and it is said here rather than left for you to notice in the delivered package.

What stays in English

The workspace your team logs into. The console, meaning settings, the review screens and the course list, is in English regardless of what language a course is produced in. There is no interface translation.

This is a deliberate split, not an oversight. The people using the console are a handful of named colleagues at your organisation; the people watching the course may be thousands, and they are the ones whose language matters.

Right-to-left languages

Arabic is supported and produces right-to-left output end to end:

• Slides are laid out right-to-left, with a script-appropriate font, and the text is shaped correctly rather than rendered as disconnected letterforms.
• Where an avatar presents a lesson, its presenter card mirrors too, so the layout does not read as if it were designed for another language and then reversed by accident.
• The course opening is produced in Arabic.

Narration is language-agnostic: the voice model speaks the target language. Where an avatar presents a lesson, its lip sync is driven by that same audio, so it is language-agnostic too.

Malayalam is written left to right, so none of that applies to it. What it needed instead was a font carrying Malayalam glyphs, which the render engine now loads.

What is available today

English, Arabic, Malayalam, Spanish, French and Hindi are the languages the video pipeline produces today.

Spanish was enabled on 1 September 2026. Its two outstanding measurements, the speech rate and the transcription floor, had landed on 9 August. What remained was a decision, and until it was made the console labelled Spanish “document only, no video yet” and refused the render outright rather than half-serving it. That state is still a supported one; no language is in it today.

Arabic, French and Hindi are priced on a speech rate that has not been measured on the current narration engine, and the bill may be wrong until the first course renders. Course length is planned from words a minute. On the engine that narrates today that figure has been measured for English and Spanish; Arabic’s was measured on the previous engine and not yet on this one, and French and Hindi have none, so length for those three falls back to the current engine’s English figure meanwhile. This is not hypothetical: the same gap in Arabic on the previous engine meant 140 words a minute was assumed where 97 were delivered, and a customer was billed for 67 minutes against 97 produced. The first render of an Arabic, French or Hindi course reports the true rate and the price is corrected against it. If you are commissioning any of the three, ask for the rate to be measured before the course rather than after it.

French and Hindi were added on the same day. Hindi needed two things that a language in a new script always needs, and both were fixed rather than waived: a font carrying Devanagari glyphs, and the danda (।) recognised as a sentence ending. Without the second, Hindi captions would have run one sentence into the next, because Hindi does not use the full stop at all, so every cue would have carried the fault, not just questions.

Adding a language is a small, well-defined change rather than a new subsystem: layout direction, font, voice behaviour and the writing rules all key off one definition per language. But a language is not supported until it has actually been rendered and checked end to end, and we will tell you plainly which list yours is on before you commit to anything.

If you need a language that is not on the list, ask. It is a reasonable thing to ask for and it is a reasonable thing for us to be honest about the timing of.

A note on quality gates

Some of the automated checks in layer 4 test for English machine-writing tells: particular stock phrases, and a punctuation habit. Those do not transfer to another script and are not applied where they are meaningless. The requirement that the writing sound native is carried in the generation rules for that language instead.

The technical layer, motion and design, and package completeness are language-independent and apply to every course in every language.

Four things are not language-independent. Each is named here rather than left for you to find, and each is also written into the QC report your reviewer reads.

Narration is only scored against the script in the languages whose transcription floor has been measured. The narration is transcribed back from the rendered audio and compared with the script you approved. A similarity percentage only means something against a floor measured for that language, on the transcriber that language is actually transcribed with, and the floors measured so far are English and Spanish. This is not a rule about the alphabet a language is written in: a language nobody has measured is treated exactly as Arabic is, whatever alphabet it uses. Arabic, Malayalam, French and Hindi have no measured floor, so for those four the comparison does not run. It records a warning saying the narration was not validated and must be verified by ear. Your reviewer sees that warning and is told plainly what has not been checked. Whether the narration audio exists at all is still a hard failure if it is missing. Measuring a floor for a new language does not need a course render, so if you are commissioning a language that is not on the measured list, ask us to measure it before your course rather than after it.

Text inside generated artwork is only read in Latin script. The reader behind the rule that no text is baked into artwork recognises Latin letters, so Arabic or Malayalam lettering inside an image is not seen. The check does not report clean when that happens. It records a warning that names the language and the number of images it could not read, and asks for those images to be reviewed by eye, which the review step puts in front of you anyway. Latin lettering that leaks into a non-Latin course still fails outright, because that is a real defect either way.

Objective levels and quiz alignment are not scored outside English. Judging whether an objective is written at the level it claims, and whether a question demands that level, needs a verb table and a set of question-demand markers for the language. Neither exists for Arabic or Malayalam yet, so those two judgements are skipped and the report says so in place of a verdict. Everything that does not depend on the language is still enforced: how many objectives there are, that each carries a real condition and criterion, that every one is served by a lesson, and that every one is assessed by a question.

A Malayalam course’s stated length, and therefore its price, may be wrong. No Malayalam speech rate has been measured yet, so the length estimate falls back to the current engine’s measured English rate. If Malayalam is spoken slower or faster than that, the minutes quoted for the course are out by the same proportion, and the price is calculated from minutes. Arabic went through exactly this: it was assumed to run at 140 and measured at 97, an assumption 44% higher than the truth, and customers had been billed for minutes they did not receive before it was caught. The course-length check reports the true figure on the first Malayalam render, and that measurement is what replaces the fallback. Until a Malayalam course has been rendered, treat its quoted length and its price as provisional and ask us to confirm both against the finished runtime.

Producing the same course twice

A course in a second language is a second production run, not a translation pass over the first. That is more work and it is also a better result: the lessons are authored in the target language rather than transliterated into it.