Languages
Spelo is built on OpenAI’s Realtime API, which natively supports multilingual input and output. You set the primary language in your site config; the AI can still understand and respond in other languages when visitors switch mid-conversation.
Supported languages
The Realtime API works well in these languages. The list grows as OpenAI improves the model.
| Language | Code | Quality |
|---|---|---|
| English | en | Excellent |
| Spanish | es | Excellent |
| French | fr | Excellent |
| German | de | Excellent |
| Italian | it | Excellent |
| Portuguese (BR) | pt | Excellent |
| Dutch | nl | Very good |
| Japanese | ja | Very good |
| Korean | ko | Very good |
| Mandarin Chinese | zh | Good |
| Hindi | hi | Good |
| Arabic | ar | Good |
| Turkish | tr | Good |
| Polish | pl | Good |
| Russian | ru | Good |
| Swedish, Norwegian, Danish, Finnish | sv, no, da, fi | Good |
“Good” means: reliable transcription and coherent responses, occasional mispronunciations. “Excellent” means: indistinguishable from a native speaker in casual testing.
Configuration
Dashboard → Voice → Language → pick one. This sets:
- The language hint passed to OpenAI at session start
- The default greeting language
- The formatting of dates and numbers (“$1,234.56” vs “1.234,56 €“)
Mixed-language sessions
Even with English as primary, visitors can speak Spanish and the AI will respond in Spanish. It’s a mid-session switch, triggered by the model itself.
If you want strict single-language, add to your custom instructions:
Always respond in French, regardless of the visitor's language.If the visitor speaks English, acknowledge in French and continue in French.Voice + language compatibility
All six OpenAI voices (alloy, echo, fable, onyx, nova, shimmer) support every listed language. Accent varies:
alloy,echo,onyx— sound more neutral / North Americanfable— slight British English inflectionnova,shimmer— sound more North American
For non-English sites, test each voice. Some sound more native than others depending on the target language.
Pronunciations across languages
The pronunciation dictionary works in every language. Write the say_as in the target language’s phonetics:
| word | say_as (for French audience) |
|---|---|
Hodos360 | HO-doze trois-cent-soixante |
Wi-Fi | wee-fee |
Data in another language
If your data is in one language but you want the AI to respond in another (e.g. English product descriptions but Spanish-speaking visitors), add to custom instructions:
Our product data is in English. When speaking Spanish, translate productnames and descriptions naturally as you read them. Keep SKU codes in theiroriginal English form.The model handles this well for mainstream product data.
Transcription accuracy
OpenAI’s Realtime uses Whisper under the hood for transcription. Whisper excels on clean audio; noisy or fast speech can cause errors. If you run multi-language sites in acoustically noisy environments (car dealerships, coffee shops), expect slightly worse transcription.
Not yet supported
- Simultaneous multi-language interfaces (one voice + caption in another language) — not yet. Pick one language per site.
- Sign language — no plans.
- Language auto-detection at first utterance — the model does this decently but we recommend setting
languageexplicitly for consistency.
See also
- Templates — industry templates ship with neutral language
- Pronunciation dictionary — works in every language
- Voices — accent comparison