Registered is not supported
The registry holds 57 languages. That number describes what the platform can represent — it is not a capability claim, and the API says so on every response.
Each language carries a separate state per capability, because a language can be perfectly collectable and completely unpublishable:
| State | Means |
|---|---|
| Registered | The platform can represent it. Nobody has claimed it works. |
| Provider supported | A supplier publishes support. VoiceInsights has not measured it. |
| VIA validated | VoiceInsights measured it and recorded the evidence. |
| Not supported | It will not work, and the platform says so rather than failing quietly. |
The five publication languages
Institutional editions ship in English, French, Portuguese, Kiswahili and Arabic. A report cannot be published in a language outside that set — the request is refused, not approximated.
Arabic and right-to-left
Arabic is right-to-left throughout, derived from the registry rather than from a setting anyone can get wrong. PDFs embed Unicode fonts with contextual shaping.
What is not validated
Speech-to-text has not been benchmarked on African speech. A supplier publishes a list of languages its model handles, and the platform records that as a supplier claim. No accuracy figure exists to quote, and asking the platform whether it can offer speech-to-text in strict mode returns a refusal.
There is also no automatic language detection. The language attached to a transcript is the language the caller declared.
Translation is governed, not generated
The platform does not translate. It governs translations produced elsewhere: provenance, source hash, staleness and review state. A translated edition must reference a governed translation artefact, or it cannot be created.