AreaAudit

Which languages each generator says it supports, and where

One page per tool, three language columns on each

Each tool page carries the same three columns, filled only from what the vendor publishes. A column the vendor never addresses is recorded as unstated rather than as a refusal. As of 2026-09-12.

What a tool page records, and what it refuses to inferA tool page is filled from the vendor's own pages. A named list is recorded as a list. A count without a list is recorded as a count, with the missing list noted. A page that never addresses a column leaves it unstated, and a localised marketing site is not treated as evidence about output.RecordedNot recordedA named listGoes in as published—A count, no listCount in, absence notedThe list is not inventedColumn unaddressedMarked unstatedNot read as a refusalLocalised websiteInterface column onlySays nothing about outputSame vendor page, four different readings
Fig. 1 The last row is the one that keeps the atlas honest: a translated marketing page says nothing about what the model will speak.

Three columns sounds like excessive precision until the first tool is read. Interface languages are usually listed somewhere obvious. Output languages are often described in a sentence with no list attached. Voice inventories are the rarest of the three, and are the one a drama production actually needs.

Every entry prints the address and the day it was read, so a claim that has since grown can be re-checked rather than argued over. Where a vendor publishes a count without a list, the count goes in and the missing list is noted beside it.

Twelve tools are covered at this depth today. The register carries the shorter form for every tool checked, including those whose pages produced nothing worth a page of their own.

1The tools

  • Kling AI — The 20 site languages Kling AI declares, the five voice languages it counts without naming, and the column it leaves unpublished.
  • SceneMixer — The 15 languages SceneMixer publishes across interface, working documents and spoken dialogue, and the 15 site alternates counted behind them.
  • CapCut — The 27 locale-tagged site alternates CapCut declares, the subtitle count it prints instead of a voice list, and the two columns left empty.
  • D-ID — Two site alternates, English and German, against a published claim of more than 120 spoken languages with no list attached to it.
  • Fliki — A named list page counting 91 languages and 117 dialects, feature pages advertising 80+ and 100+, and a site declaring no alternates.
  • HeyGen — Two named help-centre language lists with no total on either, a marketing count of 175+, and a stated ceiling of ten languages per job.
  • Higgsfield — Five same-host site alternates, one of them opened and found genuinely translated, and two columns with nothing published in them at all.
  • PixVerse — Thirteen same-host site alternates, and a language claim that is about text drawn inside the frame rather than about dialogue or documents.
  • Rask AI — Sixteen declared alternates served almost entirely from subdomains, one of them carrying a code that belongs to a different language.
  • Synthesia — A 143-row named table for the translation feature, a headline claiming 160+, a selector reading 140+, and nine site alternates.
  • VEED — Two dubbing lists rather than one: 72 languages it says it can detect in a file and 29 it says it can dub to, with no total on either.
  • Vidnoz — Twelve declared alternates served mostly from subdomains, and one page carrying two different spoken-language counts, 100+ and 140+.

A tool earns a page of its own when its pages carry enough to fill more than one column.

The rest of the atlas: Columns, Entries, By language, Side by side, Markets, Output, Questions, Terms, Learn, Data. How a list earns a column is set out on the counting page.