Multilingual conversations

How AI Avatars handle multiple languages. Single language vs Multilingual mode, voice naturalness tradeoffs, what works and what does not.

Home › Help Center › AI Avatars › Multilingual conversations

Your avatar can speak many languages — but the experience varies depending on which voice was paired with the Replica, whether you pick a single language or multilingual mode, and what you write in the System Prompt. This guide explains the tradeoffs.

Single language vs Multilingual

Each persona has a Language setting. You have two paths:

Pick a single language (English, Spanish, French, etc.) when your audience is uniformly one language. The avatar will reply in that language even if a visitor speaks in a different one. Best for branded experiences on country-specific sites. Pick Multilingual when your audience may switch languages mid-conversation, or when you don't know which language they'll start in. The avatar matches whatever language the visitor uses on a turn-by-turn basis.

How voice naturalness works

The voice paired with the Replica was trained on a specific language. A voice trained on English speaking Spanish will sound understandable but slightly accented. A voice trained on Spanish speaking English will sound the same way in reverse.

If your audience is primarily one non-English language, train a Replica recording yourself speaking that language. The result feels markedly more native than a stock English Replica speaking the same language.

What the System Prompt should say

If you've set the Language to a specific language, write the System Prompt in that language too. The avatar reads the prompt to figure out personality — a Spanish-speaking avatar driven by an English-written prompt sometimes drops into English when discussing rules or instructions.

If you're using Multilingual, you can write the prompt in any major language. The avatar will translate the personality across the languages it speaks.

Knowledge Base and language

Documents in the Knowledge Base can be in any language. The avatar will quote them in the language they were written in. If you need a fact available in two languages, attach both translated versions and the avatar will pick the one matching the conversation language.

What won't work as well

Mixed-language sentences. "Tell me about el precio de envío to Mexico" — the avatar handles this but the voice may sound jerky as it shifts mid-phrase.
Tonal languages with stock voices. Mandarin, Cantonese, Vietnamese sound much better with a Replica trained on those languages than with a stock English-trained one.
Highly localized slang. The avatar speaks standard formal versions of each language. Regional slang (Quebec French, Andalusian Spanish, Aussie English) gets understood but not produced.

If you serve multiple language audiences, create separate personas per language rather than one Multilingual persona. The single-language version always sounds more natural, and your System Prompt can be written natively for each audience.