Where AI video tools say they work, and in which languages

AreaAudit registers what AI video vendors publish about languages, in three separate columns: the site, the output and the voices. Nine of the 18 tools registered declare a site language list, two more declare a single same-host alternate, nine publish something about voices, and one states what language its documents come out in. As of 2026-09-22.

Inclusion rule. Tools with a public English site selling AI video generation or AI video dubbing. Site languages count hreflang alternates served from the host being read; x-default and alternates pointing at another host are not counted, for every vendor alike. Order. Alphabetical by tool name.

1What is registered here

One row per tool, three columns, the same rule applied to every vendor. Site languages come from the alternates a vendor declares on its own domain, which is the one measure available for every vendor and checkable by anyone in a browser. Output and voice languages come only from what the vendor publishes about its own product.

The columns are kept apart because they measure different systems and rarely move together. A site in twenty languages tells a buyer nothing about whether the dialogue can be generated in any of them, and that is the question a production actually has to answer before committing to a tool.

2Start here

• The language register — 18 tools, three columns, one table.
• Kling AI — the second-longest declared site list, and a voice count without names.
• SceneMixer — 15 site alternates, and the only tool publishing all three columns.
• Document languages — what the outlines and storyboard scripts come out in.
• Dialogue languages — what the characters are spoken in, and by whose count.
• How counted — the rule, and why nothing is inferred.

3Questions the register answers

• A translated website is not a translated product — the two counts do not match.
• Why a count is not a list — five unnamed answers nothing about Portuguese.
• The tools that declare nothing — four publish nothing in any column.
• Speaking the line versus dubbing it — where the decision sits in the pipeline.
• Why alternates pointing elsewhere are not counted — one rule, applied to everyone.

4By market

The declared lists, cut by region. Each page says which codes are declared by whom, what nobody declares, and why the spoken-language question cannot be answered per language from what is published.

• East Asia — Japanese in eight of the nine declared lists, Korean in seven; one vendor adds Cantonese for dialogue only.
• Southeast Asia — Indonesian declared by four vendors, and Filipino under two different codes.
• Europe — German in eight of the nine declared lists; five languages declared by one vendor alone.
• Middle East — Arabic declared by three of the nine, right-to-left addressed by none.
• Latin America — Spanish and Portuguese declared by seven each; three vendors name pt-BR.
• South Asia — one language, one vendor, for a region that size.

5Explainers

How language support is structured in this category, and how the one checkable piece of it works. Counts for individual vendors live in the register, each with the rule used and the date it was read.

• Three layers — interface, working documents and spoken dialogue are three separate settings.
• How hreflang works — what a declared language alternate is, and what it is not evidence of.
• Subtitles and localisation — three layers to change a market at, and the trap that belongs to each.

6What a blank cell means

It means the vendor publishes no list, not that the capability is absent. Most of these products generate dialogue in more than one language; almost none says which. Recording that absence plainly is the point of the register, because a capability a buyer cannot verify before paying is not one they can plan a series around.

The same restraint applies to the counts that are filled in. A declared site language is evidence that a vendor has translated its marketing, and nothing more; a voice count without names answers no question about a particular market. Both are recorded as exactly what they are, with the page they were read from and the date, so a reader can weigh them rather than take a summary on trust.

AreaAudit registers published vendor statements. It does not test products and it does not rank them.

7Where to start reading

The register sections are built from vendor pages and dated; the rest is explainers. Start with a market if you know where you are selling, with a tool if you already have one in mind.

• Data — Every recorded language claim as one comma-separated file
• Entries — A page for each register cell worth reading on its own…
• By language — Turn the register sideways…
• Side by side — Pairs of vendors that answer one register column in opposite ways
• Terms — Definitions for the words vendors and this atlas use about languages…
• Columns — The three columns every recorded language claim is filed under, why they are never merged…
• Markets — A page per market…
• Tools — A page for each generator…
• Output — What generators publish about the language of the documents they produce…
• Questions — Questions about which languages AI video tools name, answered from vendor pages…
• Learn — Explainers on reading a vendor's language claims…