An AZ8 editorial hands-on test
Getting AI video to speak Japanese: what 7 models did
More and more models generate video with sound, so a character can now speak the Japanese lines you write in your prompt. We generated the same lines with 7 models and checked how each one actually sounds.
Last updated AZ8 Editorial Team
Which model should I use for Japanese dialogue?
In our test, Seedance 2.5 and Kling 4.0 delivered both lines word for word. Seedance 2.0, Wan 3.0, and Gemini Omni Flash also followed the script, apart from spelling differences in the transcript and one extra sound at the end of a line. Every model speaks Japanese as long as you wrap the lines in Japanese quotation marks 「」.
- All 7 models spoke the scripted lines in Japanese
- Seedance 2.5 and Kling 4.0 made zero errors, and even the station announcement came out in Japanese
- Each model was generated once, with no cherry-picking
The same lines, generated by 7 models
We generated the prompt below once with each model and transcribed the audio with Whisper (medium model). The character error rate is the number of characters that differ from the scripted lines, divided by the length of the lines.
Prompt
夕方の駅のホームで、制服姿の女子高生が友達に振り向いて、日本語ではっきりと「また明日ね、遅れないでよ」と言う。友達は笑って「わかってるって」と答える。電車の到着音と駅のアナウンス。
Scripted line
- 「また明日ね、遅れないでよ」
- 「わかってるって」
| Model | Transcript (Whisper) | Character error rate |
|---|---|---|
| Seedance 2.5 | また明日ね 遅れないでよ わかってるって | 0% |
| Kling 4.0 | また明日ね 遅れないでよわかってるって | 0% |
| Seedance 2.0 | また明日ね、遅れないでよ 分かってるって | 6% |
| Wan 3.0 | また明日ね。送れないでよ。 わかってるって。 | 6% |
| Gemini Omni Flash | また明日ね遅れないでよ わかってるってよ | 6% |
| MiniMax H3 | また明日ねおこれないでよ わかってるって | 11% |
| Veo 3.1 Lite | これ、また明日ね。遅れないでよ。分かってるって。 | 17% |
Seedance 2.5
- 16:9
- 720p
- 8s
Spoke both lines word for word. The station announcement we specified in the prompt, 「まもなく1番線に電車が参ります」, also came out in Japanese (excluded from the error rate).
Kling 4.0
- 16:9
- 720p
- 8s
Spoke both lines word for word. A station announcement, 「ご注意ください、電車が参ります」, also plays in Japanese at the start (excluded from the error rate). The two characters facing each other look natural, too.
Seedance 2.0
- 16:9
- 720p
- 8s
The lines match the script. The only difference is that 「わかって」 was transcribed as 「分かって」, a spelling variant that sounds the same.
Wan 3.0
- 16:9
- 720p
- 8s
「遅れないで」 was transcribed as 「送れないで」, but both are read "okurenaide." As audio, it matches the script.
Gemini Omni Flash
- 16:9
- 720p
- 8s
Spoke both lines almost exactly. The only difference is an extra 「よ」 at the end of a line.
MiniMax H3
- 16:9
- 768p
- 8s
「わかってるって」 is exact. 「遅れないで」 sounded like 「おこれないで」. It is the only one of the 7 models that outputs a wide 1344×768 frame.
Veo 3.1 Lite
- 16:9
- 720p
- 8s
Both lines are clearly audible Japanese. It added the word 「これ」 before the lines. It is the cheapest of the 7 models to try, in credits.
How to write prompts that keep Japanese dialogue intact
- 01
Wrap the lines in 「」
Separating the dialogue from the description tells the model exactly what to say out loud. Our test prompt also wraps both lines in 「」.
女子高生が「また明日ね、遅れないでよ」と言う。
- 02
State "in Japanese" explicitly
Naming the language keeps the lines from being swapped for English or another language. Adding how to speak, such as "clearly," makes the lines easier to hear.
日本語ではっきりと話す。
- 03
Keep each shot's lines short
A 5 to 10 second video fits about two short lines. For a longer conversation, generate a separate shot for each line and join them.
1 to 2 sentences per shot
Which model for which goal
Seedance 2.5
When line accuracy comes first
One of the two models that spoke without a single error in our test. It generates up to 30 seconds at once, so a scene with conversation can be made without cutting.
See Seedance 2.5 detailsMiniMax H3
Official Japanese support
The developer states that dialogue stays stable in 11 languages, including Japanese. It also supports 21:9 widescreen and 2K.
See MiniMax H3 detailsKling 4.0
Cinematic scenes with dialogue
The other model that spoke without a single error in our test. Up to 30 seconds, with 21:9 widescreen support.
See Kling 4.0 detailsSeedance 2.0
Commercials and dramas finished in 4K
Its lines also matched the script in our test. It offers 4K on AZ8, so close-ups of a speaking character stay sharp.
See Seedance 2.0 detailsJapanese dialogue FAQ
Do AI video generators support Japanese?
All 7 models we tested spoke the Japanese lines from the prompt in Japanese. MiniMax H3 is the one whose developer explicitly states Japanese support. The other models also accept Japanese prompts as is.
How do I make an AI video character speak Japanese?
Choose a model that generates audio, and write the lines in your prompt wrapped in 「」. State the language with "in Japanese" (日本語で) and describe how to speak, and the result is more likely to match what you intended.
Will I get the same result every time?
Even with the same prompt, results vary from one generation to the next. These results come from a single generation per model, with no cherry-picking. If a result is off, generating again with the same prompt sometimes improves it.
How is the error rate calculated?
We transcribed the audio of each video with Whisper, removed punctuation and spaces, and divided the character differences from the scripted lines (edit distance) by the length of the lines. A kanji versus kana spelling difference also counts as one character.
Can I control the voice or delivery?
Delivery such as "in a small voice" or "cheerfully" tends to come through when you write it in the prompt. To get closer to a specific voice, use a model that takes audio as a reference, such as Seedance 2.5, Seedance 2.0, Wan 3.0, or MiniMax H3.
Can I use the videos commercially?
Under the AZ8 Terms of Service, you may use the videos you generate commercially as long as you comply with the law, respect third-party rights, and follow any model-specific conditions shown before generation. Avoid generating content that imitates a real person's voice or face.
Try it right in your browser.
Signing up is free. Generating videos uses credits.