Skip to main content
Each endpoint has one model behind it, named after its task. You never pick a model in a request: call the endpoint and the API uses the model that serves it. The list is also live at GET /v1/models, which needs no key.
speech-recognitionBetaWrites down what was said. Hausa, Zarma and French.POST /v1/transcribe · scope transcribespeech-synthesisBetaSpeaks text aloud with a chosen voice. Hausa, Zarma and French.POST /v1/speech · scope speechtranslationBetaHausa ⇄ French, from text or from speech, with the result spoken if you ask.POST /v1/translate · scope translatevoice-agentResearchHolds a spoken conversation in Hausa.POST /v1/converse · scope converse

​
Languages by model

Translation works between Hausa and French only, in both directions. Zarma translation is not in the API yet.

​
What the status means

Ready to build on, with real users, while we keep improving quality. The endpoint and its fields will not change without notice in the changelog. Quality is measured and improving, so test with your own sentences and listeners before you rely on it for anything that matters.
A preview of work in progress. Use it to explore and give us feedback, not for something people depend on. It can change, be slower, or answer in ways we are still learning about. New keys do not include its scope unless you tick it.

​
Older names

On 7 November 2026 the models took the names of their tasks. If you see the old names in an older integration or message, they are the same models: Requests never name a model, so nothing in your code needs to change.

​
Things to know

  • Transcripts come in lower case, without punctuation.
  • Speech comes back as WAV by default; ask for opus for the smallest files (audio formats).
  • The first request after a quiet spell is slower while the model wakes up. Allow up to a couple of minutes on that first call, and keep a generous timeout.
  • Where the work happens. speech-recognition and speech-synthesis run on Namu’s own models. translation and the voice-agent’s replies are written by Google Gemini through OpenRouter, which receives the text they work on.
  • Native review. The Hausa and Zarma examples in these docs have not all been reviewed by native speakers yet. If you see a mistake, tell us.