AI Video With Non-English Speech - What Works (My Test)

Most AI video models with native audio speak English well, but other languages are hit or miss, and you only find out by testing. On 28 September 2026 I gave eleven video models the same hard Polish sentence: MiniMax H3 Max, Grok Imagine Video 1.5, Gemini Omni Flash 1.1 and Veo 3.1 got it right (Veo almost), MiniMax H3 and FLUX.3 Video missed one word, and Kling 3.0 Pro got 10 of 12 words wrong. On Jantaris the simplest starting point for non-English lines is MiniMax H3 (from $0.161 per second, stereo audio always included). Below: the full results, how to test any language yourself, prompt rules, cost per clip and when to record a voice-over instead.

Why non-English speech fails

Video models learn speech from their training data, and most of it is English or Chinese. Model makers publish short official lists of supported speech languages. For example, Kling 3.0's list is Chinese, English, Japanese, Korean and Spanish. If your language is not on the list, the model may still try, but with the phonetics of a language it knows better.

Two failure modes I saw:

• Foreign accent on correct words. The words are recognizable, but every consonant cluster sounds American.
• Garbled words. Syllables merge, endings disappear, numbers turn into nonsense. Speech recognition cannot even identify the language.

My test: one Polish tongue twister, eleven models

Polish is a good stress test: consonant clusters, nasal vowels and inflected numbers. Every model got the same prompt: a woman at a kitchen table looks into the camera and says "W Szczebrzeszynie chrząszcz brzmi w trzcinie, a ja płacę trzydzieści dwa złote" (a classic Polish tongue twister plus a price). Settings: 5 seconds, 16:9, 720p or the nearest tier, audio on, one take per model, no rerolls.

I checked the audio with faster-whisper speech recognition in two sizes (large-v3 and medium) and counted misrecognized words out of 12. The clips were generated through other API providers, not Jantaris, so treat the results as a signal about the model, not a guarantee. Full write-ups (in Polish) are in my video model ranking and Kling in Polish test.

What I take from it:

• MiniMax H3 is the safe pick on Jantaris. One slip at the start of the sentence, the rest clean, and audio comes with every clip at no extra cost.
• Grok Imagine Video 1.5 and FLUX.3 Video also held Polish pronunciation well. On Jantaris, Grok needs a start image.
• H3 Max scored best, with a caveat. The provider I tested on described H3 Max differently from the Jantaris catalog, and the Jantaris description does not list an audio track. Run one test clip before planning a batch.
• Veo 3.1 has no audio switch on Jantaris, so I do not carry its good result over one to one.
• I do not use Kling or Seedance for Polish lines. Great pictures, broken words.

The three-language clip was the most interesting result. Kling read Polish and German with American phonetics while its English was clean; MiniMax H3 read English and German with a strong Polish accent. Each model seems to have one "native" voice and keeps its accent when the language switches. That is my reading of two clips, not a rule. The practical lesson: if you need English, Kling is a good choice; if you need Polish, it is not.

How to test your language in ten minutes

You do not need to trust my results for your language. The same method works for German, Portuguese or Hindi:

1. Write one hard sentence. Include the sounds your language is known for, a number and a proper noun.
2. Generate 5 seconds per model at the lowest resolution, same prompt, one take each.
3. Transcribe the audio with Whisper or any speech-to-text tool, with the language set explicitly.
4. Count wrong words against the original sentence. Also run language detection without a hint: if it guesses the wrong language, viewers will hear it too.
5. Listen once for accent. Transcription can forgive an accent that a native speaker will not.

At the lowest tiers this costs a few clips, for example 5 × from $0.161 per second for H3 at 768P and 5 × from $0.0777 per second for Wan 3.0 at 480P.

Prompt rules for non-English lines

I describe the scene in English and keep the spoken line in the target language, in quotes, with the language named:

Medium close-up of a man in his forties at a workshop bench, looking into the camera.
He says clearly and calmly in German: "Guten Tag, heute zeige ich, wie man eine Dichtung wechselt."
Static camera, natural daylight, quiet workshop ambience, no music, no subtitles, no on-screen text.

• Spell out numbers. "thirty-two" in the target language, not "32", so the model does not have to guess the inflection.
• One short sentence per 4-5 seconds. All eleven models fit my 12-word sentence into 4-5 s; longer text gets rushed.
• Add "no subtitles". I always include it so the model does not print the line on screen.
• Describe the voice. Age, tone and pace. Without it you get a random voice.

Generating the clip on Jantaris

In Jantaris Studio open the Video tab, choose MiniMax H3, paste the prompt, use 768P for drafts and 2K for the final. Through the API:

curl https://jantaris.com/v1/generate \
  -H "Authorization: Bearer $JANTARIS_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "minimax-h3",
    "input": {
      "prompt": "Medium close-up of a man in his forties at a workshop bench, looking into the camera. He says clearly and calmly in German: \"Guten Tag, heute zeige ich, wie man eine Dichtung wechselt.\" Static camera, natural daylight, no music, no subtitles.",
      "aspect_ratio": "16:9",
      "resolution": "768P",
      "duration": 6
    }
  }'

In text-only mode H3 needs an explicit aspect ratio (adaptive is rejected). Starting from a photo of a person, send it in image_url and the ratio comes from the image. Fetch the result with GET /v1/tasks/{id}. H3 image quality at 2K from my August test:

Cost of a clip with speech

Seconds times rate, times attempts. With speech I plan 2-3 attempts, because even a good model occasionally swallows a word:

"From" rates are for the lowest resolution; H3 at 2K costs more. Failed generations cost $0, and the balance starts at $10 with no expiry. Other combinations are in the AI cost calculator.

When to record a voice-over instead

Generated speech works for short lines by a character on screen. Otherwise I use a separate voice:

• Off-screen narration. Tutorials, product films, ads with a narrator. I generate silent video (for example Kling 3.0 without audio at 720p, from $0.169 per second) and add a recorded or synthesized voice in the edit. Lip sync does not matter.
• Long scripts. A minute of speech is several clips, and each may get a different voice. One narrator sounds consistent.
• Brand names and jargon. A model may mispronounce your product name in every take. A recording gets it right once.
• English line plus subtitles. If the audience accepts it, an English line from Kling with subtitles in your language is often cleaner than a garbled native line.

Phone recordings benefit from cleanup, for example with Voice Cleaner (see AI voice noise removal). For a character moving to your own recording, Wan 2.7 Video takes an image and an audio file together (image_url + audio_url); I have not tested how closely it matches mouth shapes. Related comparisons: MiniMax H3 vs Kling 3.0 and Kling 3.0 vs Veo 3.1. For speech from a single photo, see AI talking photo.

FAQ

Which AI video generator speaks languages other than English?

In my Polish test MiniMax H3 Max, Grok Imagine Video 1.5 and Gemini Omni Flash 1.1 were error-free, Veo 3.1 nearly so, and MiniMax H3 and FLUX.3 Video missed one word. Results differ per language, so run a short test in yours.

Does Kling speak languages other than English?

Kling 3.0 officially lists Chinese, English, Japanese, Korean and Spanish. In my Polish test the Pro tier got 10 of 12 words wrong, while its English was clean.

How do I prompt a video model to speak a specific language?

Describe the scene in English, put the line in the target language in quotes, add "says clearly in German" (or your language), spell out numbers and add "no subtitles". Keep it to one short sentence per 4-5 seconds.

How can I check the speech quality of an AI video?

Transcribe the audio with Whisper or another speech-to-text tool with the language set, count wrong words against your script, and run language detection without a hint. Then listen once for accent.

How much does a clip with speech cost?

Seconds times the model's rate. Six seconds of MiniMax H3 at 768P is 6 × from $0.161 per second, audio included; three attempts triple it. Failed generations cost $0.

Should I use generated speech or a voice-over?

Generated speech for short on-screen lines, a separate voice-over for narration, long scripts and brand names. A silent clip plus a recorded voice is the most predictable option in any language.