I build an AI tool that writes wedding speeches in Japanese, French, Spanish and English. The hardest part was not the generation. It was the etiquette.
Japanese weddings have a concept called imikotoba (忌み言葉) — words you avoid because they suggest separation, ending, or repetition. Say 切る ("to cut") in a wedding speech and the room notices. Say 別れる ("to part") and someone's grandmother quietly disapproves.
The problem
Every LLM will happily write a fluent Japanese wedding speech. Almost none of them know that 終わる ("to end") is a landmine in that context, or that 戻る ("to return") suggests going back to a previous state — not something you want in a blessing.
Prompting helps. It does not solve it. Because:
- The list is not in the training data in a structured way
- Post-generation filtering needs an actual list
- "Avoid these words" is useless without knowing what to say instead
So I extracted the list from my own content library and published it.
What's in the dataset
5 categories, 19 word pairs, plain JSON:
-
wakare— separation (切る, 別れる, 離れる) -
owari— ending or returning (終わる/終わり, 戻る, 帰る) -
kasane— repeated words (たびたび, ますます, 重ね重ね) -
suuji— the numbers 4 and 9 (死 / 苦) -
sonota— everything else
Each pair has avoid and instead. It's licensed CC-BY-4.0, so commercial use is fine.
Repo: https://github.com/johnsiebyron-stack/imikotoba
It's also on npm as imikotoba if you'd rather import it. (speechkobo.com is the product this came from.)
The part that surprised me
I added a small helper that splits variants on slashes and middots. Turned out this matters a lot.
The raw data has entries like 終わる/終わり and 倒れる・落ちる・枯れる. If you naively match the whole string, you match nothing. My first version did exactly that and silently missed every variant.
// Before: matches nothing
text.includes("終わる/終わり") // false
// After: split into base forms
baseForms("終わる/終わり") // ["終わる", "終わり"]
That bug would have shipped. The test caught it, which is the only reason I'm writing this post instead of quietly failing in production.
Tradeoff: false positives
I use substring matching on purpose. 区切る contains 切る, and find() will flag it.
That's a deliberate choice. In a speech checker, a false positive costs you two seconds. A false negative costs you the moment. But if you need precision, you'll have to filter — and the README says so explicitly instead of pretending the library is smarter than it is.
What I'd tell anyone building in a non-English market
- The etiquette layer is not optional. Generic generation gets you 80% of the way. The last 20% is local knowledge that isn't in any API.
- Publish the data separately from the product. It forces you to make it explicit, and it makes your own prompt shorter.
- Count characters, not words. Japanese speech timing is ~300 characters per minute. English is ~150 words per minute. Same 3 minutes, completely different target numbers.
Feedback welcome
If you know Japanese wedding customs better than I do, the dataset is wrong somewhere and I'd like to know where. Issues and PRs are open. The list is a practical one, not an authoritative source — regional and family customs differ, and I'd rather fix specifics than defend the list as-is.
Top comments (0)