AreaAudit

Which languages each generator says it supports, and where

Language questions settled by published lists alone

A question belongs here when a published list settles it. Anything needing a test render, an account or a support reply is out of scope, because a reader could not repeat the check. As of 2026-09-12.

Which language questions a published list can settleA published list settles whether a language is named, in which column it appears, and when the claim was last read. It cannot settle how good the output sounds, because judging that needs an account and a render, and a reader could not repeat the exercise.What would settle itA published listIt gets a pageNamed, in a statedcolumn, with the addressand the reading date.A renderOut of scopeQuality needs listening,and listening needs anaccount.A support replyOut of scopeAn answer in an inbox isnot something anotherreader can open.Only the left branch stays checkable
Fig. 1 The branch on the left is narrow on purpose. Everything on the right would make this site the source of the claim rather than the vendor.

The restriction bites hardest on the question asked first: how good the output is in a given language. Quality needs listening, listening needs an account, and an account makes the answer unrepeatable. What can be settled is narrower and still useful — whether a language is named at all, in which column, and since when.

Answers open with the finding and then show the rows behind it. Where the lists disagree with each other — a language named for voices but not for dialogue — the disagreement is the answer rather than something to average away.

1The questions

  • Site vs product — Two tools declare site languages.
  • Counts vs lists — One vendor publishes five voice languages without naming them, and a count answers a much smaller question than a named list does.
  • The silent six — Six of the eight tools tracked here publish no language information at all, in any of the three columns this register keeps.
  • Performing in-language — One vendor states the model performs the line in the chosen language as it renders, which is a claim about method rather than quality.
  • How counted — Site language counts here use same-domain hreflang alternates only.

Where the atlas cannot answer, the page names the sentence a vendor would have to publish for an answer to exist.

The rest of the atlas: Markets, Tools, Output, Learn, Data. How a list earns a column is set out on the counting page.