Eleven v4 claims 90+ languages: check Korean before you ship

Treat 90+ languages as a list of languages the model accepts, not as proof that Korean ad copy will sound right. ElevenLabs states the figure for Eleven v4 (model id eleven_v4), and its announcement also lists IPA phoneme support and a pronunciation dictionary. None of that replaces a listening pass by a native speaker on your own product names, numbers and prices.

The pages read give no per-language quality figure for Korean, so this post gives a QA routine rather than a verdict.

What the vendor pages state

A five-step Korean QA routine

Run this on a fixed test script before any batch render. Keep the script and settings so the next model release can be compared on identical input.

• Build a 10-line script with the hard cases: brand names in Latin letters, prices and phone numbers, dates, a loanword, and a long sentence.
• Render it with each candidate voice and settings, and keep every output.
• Have a native speaker mark each line: correct, odd stress, or wrong reading. Count the lines, do not average a feeling.
• Fix reading problems with the pronunciation dictionary or phoneme input rather than by respelling the copy.
• Re-run the same script after any model or voice change.

What Sume checks, and what it does not

Sume does not ship Eleven v4; its TTS Router serves Cartesia Sonic models only. The language handling it does add sits before the audio is made. Each stored voice has a primary language, set when the voice is created. If the language on your request disagrees with it, the API returns HTTP 409 tts_voice_language_mismatch before any job is created or charge made. You confirm by retrying the same request with confirm_language_mismatch: true.

For script-bound jobs, the job also carries a receipt with the submitted text hash. The Sume TTS reference is explicit that this proves the submitted text, not the pronunciation, so listening QA stays a separate step on any engine.

curl -X POST https://api.sume.com/v1/tts-router/generate \
  -H "Authorization: Bearer $SUME_API_KEY" \
  -H "Content-Type: application/json" \
  -H "Idempotency-Key: ko-check-001" \
  -d '{
    "model": "sonic-3.6",
    "transcript": "Test line for review.",
    "voice": { "id": "YOUR_VOICE_UUID" },
    "language": "ko"
  }'

The request above omits confirm_language_mismatch on purpose. If the voice's primary language is not Korean, the API answers 409; add confirm_language_mismatch: true to a retry only after you have asked a person; the guard exists so a wrong voice does not reach a paid render silently. Drop the field when the voice language already matches.