Grok TTS language "auto" vs Sume's explicit language field

Does Sume detect the language of a TTS script like Grok's "auto"?

No. xAI lists "auto" among its language options, so the service works out the language of your text. Sume TTS 1.0 asks you to name it: the language field takes a BCP-47 or ISO-639 code such as ko, ja or en, and the contract says to set it for every non-English transcript.

If you leave it out, the OpenAPI description says the request defaults to English at the provider. Sume adds one fallback: it infers ko or ja from a transcript that is only Hangul or only kana.

What happens to a Spanish script with no language set?

It is spoken with English pronunciation rules. The words are the same, so the job completes and is billed, and you hear an accent mismatch instead of an error. That is the main risk of moving from an auto-detecting API to an explicit one.

Other languages are not in the fallback. Hindi, Arabic, French and every Latin-script language other than English need an explicit code.

How do you make your code set the language every time?

Make language a required argument in your own wrapper, take it from the content record, and send it on every request. A request whose voice has a different primary language returns HTTP 409 tts_voice_language_mismatch, or a non-error warning in the MCP tool that you confirm and retry with the same idempotency key. The check happens before any job is created or credit reserved.

curl -X POST https://api.sume.com/v1/tts-1.0/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: es-welcome-001" \
  -d '{
    "transcript": "Bienvenidos a nuestra tienda. Hoy tenemos ofertas.",
    "avatar_handle": "narrator",
    "language": "es"
  }'

How do you detect language yourself before sending?

If your pipeline receives mixed-language text, detect in your own code or in the system that wrote the copy. A translation step usually knows the target language already, so pass that value through rather than guessing from text. Mixed-language sentences are the hard case; xAI does not describe how "auto" resolves them in the page read, and Sume reads the whole request in the one language you give.

Split code-switched copy into one job per language run, set each language, and join the results with timeline audio. It costs a few more jobs and removes the guesswork.

What are the limits of this advice?

The xAI page lists 20 languages plus auto in the part we read. Sume does not publish a language list in the TTS contract, so test each language you plan to ship with a short sample before a batch.

The related posts cover Hinglish and Odia and Urdu codes, which are the cases people most often get wrong.