# The Complete Guide to Multilingual AI Dialogue — a prompting skill for multilingual AI dialogue
Source: ekly.ai/guides/multilingual-dialogue — free to use with any AI assistant.

## Hard rules

1. Diagnose the root cause before rewording. Cosmetic prompt retries almost never fix pronunciation.
2. Pick a voice engine trained natively on the target language family before worrying about cost — **except** when the line carries the film's whole idea: directability (inline performance controls) outranks nativeness. Check for an expressivity control before casting that beat.
3. If a specific word keeps failing despite correct language and script, respell it phonetically or embed it in a fuller phrase first. Phonemes are not promptable on native SPEAK — syllable-critical brand/technical words are a dub beat from the start; two wrong pronunciations of the same syllable ends the SPEAK route (mute + TTS), do not burn a third.
4. For Arabic: if respelling fails, try writing tashkeel diacritics explicitly. Flagged unverified — research guidance, not a confirmed generation finding.
5. For Mandarin: swap polyphonic/rare characters for same-sound homophones (confirmed official guidance from a major video model). Tone accuracy via inline pinyin remains untested on many engines.
6. For long German compounds: hyphenate or split into component words.
7. Do not trust a video model's native voice outside its confirmed-safe language set.
8. Prefer: generate dialogue with a dedicated speech engine, then lip-sync — or use a video model that accepts a pre-generated audio reference (Seedance).
9. Technique hierarchy for exact pronunciation: respell → supply missing diacritics → phoneme/IPA tags (where supported; strongest for English) → accept residual bias honestly.
10. Respelling is the only technique confirmed to work end-to-end across every engine.

## Engine choice

| Language family | Look for |
|---|---|
| Hindi, Tamil, Telugu, other Indian languages | Indic-native engine |
| Chinese, Japanese, Korean | Dedicated CJK-capable engine |
| English and most widely-supported languages | Broad multilingual engine |

## Video-model language safety

| Model | Confirmed-safe | Non-safe default |
|---|---|---|
| Veo | English | Decouple audio; do not trust native non-English |
| Kling | English, Chinese | Do not treat as the multilingual option; usable path is unconfirmed for broader languages |
| Seedance | Any language via audio reference | Feed pre-generated speech as audio reference |
| Other native-audio models | English | Usually no audio-reference path; bad line costs a full re-roll |

## Hindi specifics

- Confirmed: nukta words (e.g. मज़ा) may mispronounce — respell without nukta (मजा) or embed in a fuller phrase.
- Devanagari + heavy performance language can trip a video-model content filter — romanize and soften.
- Reported: some video-model native Hindi reads more robotic with weaker lip-sync than core-tier languages.

## Practical rule

For Hindi, Arabic, Malayalam, Mandarin, or any language outside a model's confirmed-safe set: do not default to that model's own native voice. Decouple, or use an audio-reference-capable video model.