A product has three language layers, not one

Interface language is what the buttons are in. Output language is what the outlines and storyboard scripts come out in. Dialogue language is what the characters speak. A tool can be strong in one and silent about the others. As of 2026-09-12.

Inclusion rule. The three layers this register keeps apart. A layer no vendor addresses would still be listed, because an empty layer is a finding about disclosure. Order. Alphabetical by layer.

1Why one number cannot describe a product

A site translated into twenty languages can generate dialogue in one. A tool that writes outlines in eight languages might speak none of them. And a voice library covering five languages says nothing about what the settings menu is in.

These are done by different teams with different budgets and they have no reason to arrive at the same number. A single supports N languages claim therefore describes whichever layer the marketing team had in mind, and a reader cannot tell which.

2The interface layer is the cheapest to add and the easiest to see

Translating a product interface is content work: strings, a switcher, a set of alternate URLs. It signals where a vendor expects buyers to be, which makes it genuinely informative about the company.

It is also the layer most visible to somebody evaluating quickly, which is why it gets mistaken for the whole answer. Seeing your language in the switcher tells you the marketing site was translated and nothing else.

3The document layer decides whether a crew can work

Outlines, storyboard scripts, character descriptions and shot lists are handed to directors, designers and editors. If they arrive in a language the room does not read, the tool has not removed the translation step, it has moved it downstream to somebody less equipped.

This layer is rarely mentioned on product pages, and for a non-English production it is often the one that decides whether the tool is usable at all.

4The dialogue layer is a casting constraint

What the characters speak is the layer an audience experiences. It is also the hardest of the three, because it requires a voice per language, per character, that holds across a series.

Where a vendor publishes a named list here, a producer can check their market in one second. Where a vendor publishes a count, they cannot check anything at all, whatever the number is.

5What else changes besides the words

A market version is more than translated dialogue. Character names, the ethnicity written into appearance descriptions, the architecture and street furniture in a scene, the products on a shelf, and any text visible in frame all belong to a market.

A production that localises only the dialogue produces something that sounds local and looks imported. Whether a tool addresses that is a separate question from how many languages it speaks, and almost nothing published addresses it.

6The layers fail in a specific order

In practice a product adds them in order of cost: interface first, because it is strings; documents second, because it is a prompt instruction; dialogue last, because it needs voices and models that speak.

So a vendor with dialogue coverage almost certainly has the other two, while a vendor with a translated site may have neither of the others. That asymmetry is useful when reading a page that only mentions one layer: which layer it mentions tells you roughly where the product is.

7How to check a vendor's claim in five minutes

Open the site in the language you care about and see whether it is genuinely served or a machine swap of the English page. Then look for a named list rather than a count, in both the document and the dialogue layers. Then ask directly about anything that is a count.

That sequence takes minutes and settles more than an hour of reading marketing copy, because it separates the layer that is easy to translate from the layers that are not.

8Where the published answers currently sit

Across the eight tools tracked here, six publish nothing in any layer. Of the two that publish anything, SceneMixer is the one publishing all three, with 15 named languages covering interface, working documents and dialogue, plus Cantonese available for dialogue only and region presets that adapt names and appearance. Kling AI declares the wider site coverage at 20 languages and publishes a voice count of five without naming them.

Which of those matters depends entirely on whether you are buying a translated site or a translated production. The register keeps them in separate columns so the question can be asked that way round.

9Where the language counts are recorded

Counts and lists per vendor are kept in the register, each read from the vendor's own domain under one counting rule applied identically to everyone.

• SceneMixer — publishes all three layers, with 15 named languages and Cantonese for dialogue only
• Kling AI — declares 20 site languages and publishes a voice count without naming it

An explainer about method and structure, not a count. No vendor number appears here that is not also in the register. The sourced material is on the register. Related: How hreflang works, Three layers.