Multi-Language Assistants
Set up one assistant that answers callers in more than one language — detecting the language from the caller's first words and replying in that language for the rest of the call.
Screenshot ID: MULTILANG-001
Filename: MULTILANG-001_transcription-settings.png
Description: Voice tab scrolled to the Transcription section, showing the model dropdown and the Transcription Language field
Status: 🔴 Not started
Before you start
A multi-language assistant is one assistant on one phone number. It is not a menu ("press 1 for English"), and callers never have to choose. They simply speak, and the assistant follows.
Three separate things all have to support a language before your assistant can hold a conversation in it:
| Layer | What it does | What limits it |
|---|---|---|
| Understanding | Turns the caller's speech into text | The listening mode you pick (see below) |
| Reasoning | Decides what to say | Rarely a limit — handles all common languages |
| Speaking | Turns the reply back into speech | The voice you pick |
The understanding layer is almost always the one that decides which languages you can actually offer. Start there.
Choosing a listening mode
In Voice Tab → Transcription, the Transcription Model dropdown sets how your assistant listens. Two of the options matter for multi-language work, and the dropdown tags them for you: one is marked (recommended for English) and the other (recommended for multilingual). This guide calls them the English-recommended and Multilingual-recommended models.
English-recommended — best turn-taking, ten languages
This is the default for English assistants and the one to keep if you can. It has a purpose-built sense of when a caller has finished speaking, which is what makes the assistant feel quick without talking over people.
Languages: English, French, Spanish, German, Italian, Dutch, Portuguese, Russian, Hindi, Japanese.
Use it when every language you need is on that list — for example an English + French assistant.
Multilingual-recommended — many more languages, slower turns
This mode understands a much wider set of languages, including Farsi (Persian), Arabic, Hebrew, Turkish, Korean, Chinese, Polish, Ukrainian and many more.
The trade-off is real: it does not have the fast end-of-turn detection that English-recommended has. The assistant waits on silence instead, so replies come a beat later and it is more likely to talk over a caller who pauses mid-sentence. This affects every call on the assistant, including calls in your primary language.
Use it when you need a language that English-recommended does not cover.
If English + French covers your callers, stay on English-recommended. Only move to multilingual-recommended when you genuinely need a language outside its ten — the speed cost applies to all of your calls, not just the ones in the extra language.
Setting the language
Under the model dropdown is the Transcription Language field. This is a free-text field — what you type has to match the mode you chose.
| What you want | Model | Transcription Language |
|---|---|---|
| English only (fastest) | English-recommended | en |
| Two or more languages from the ten | English-recommended | multi |
| Any language set including one outside the ten | multilingual-recommended | auto |
| One specific non-English language only | multilingual-recommended | the language code, e.g. fa |
multi means "expect the caller to switch between the supported languages mid-call." auto means "work out the language and follow it."
multi only works in English-recommendedMultilingual-recommended rejects multi and will refuse to save. Use auto there instead. Likewise, English-recommended rejects language codes outside its ten — including fa — so a Farsi assistant must use multilingual-recommended.
Worked example: English + French + Farsi
A shop serving English and French-speaking customers, whose owner also takes calls in Farsi.
Farsi is not one of the ten English-recommended languages, so this assistant has to use multilingual-recommended.
- Voice Tab → Transcription → Transcription Model → the option marked (recommended for multilingual)
- Transcription Language → type
auto - Voice → pick one voice (see below) — the same voice speaks all three languages
- Instructions → add a language block (see below)
- Greeting → greet in all three, briefly
- Save, then test each language on a real call
For an English + French assistant with no third language, use English-recommended with multi instead, and keep the faster turn-taking.
Telling the assistant to mirror the caller
Detection handles the listening. You still have to tell the assistant what to do about it, or it will tend to drift back to English. Add a block like this near the top of your Instructions:
## LANGUAGE — highest priority
You are fluent in English, French and Farsi.
1. Detect the caller's language from their first sentence.
2. Reply ONLY in that language. Never mix two languages in one reply.
3. If the caller switches language mid-call, switch with them and stay switched.
4. Never announce the switch. Do not say "I can speak French" — just speak it.
5. If you truly cannot tell, use English and let the caller correct you.
Put it near the top and mark it as high priority. Instructions further down the prompt carry less weight.
A caller who hears only English often switches to English out of politeness, even when they would rather not. A short greeting in each language — "Thank you for calling. Bonjour!" — tells them they have a choice.
Choosing a voice
Your assistant has one voice, and that voice speaks every language on the call.
Most of our voices are cross-lingual: a voice built for English will still pronounce French or Spanish correctly. What carries over is the accent. An American-accented voice speaking French is perfectly understandable, but it sounds American.
How to choose:
- If one language is clearly primary, pick a voice in that language's accent and accept the accent on the others.
- If two languages matter equally, pick a voice from the smaller-audience language. Callers are usually more forgiving of an accent in the majority language.
- For a Canadian French audience, choose a Canadian French voice rather than a European French one. The difference is obvious to Québécois callers.
Cross-lingual quality varies a lot by language. It is reliable for major European languages and much less predictable for languages written in other scripts — Farsi, Arabic, Hebrew, Thai. Use the voice preview to hear your actual greeting in each language before you launch. If it sounds wrong, it will sound wrong to your callers.
If a cross-lingual voice cannot do a language acceptably, your options are a dedicated voice in that language (which then carries that accent into your other languages), or a separate assistant on its own number for that language.
When to use separate assistants instead
One multi-language assistant is usually right. Consider separate assistants on separate numbers when:
- The speed cost of multilingual-recommended is unacceptable for your main language
- No single voice sounds acceptable across all your languages
- Each language needs genuinely different instructions, hours, or booking behaviour
- You want to advertise a dedicated number to a specific community
The cost is that callers have to reach the right number, and you maintain each assistant separately.
Testing
Detection cannot be verified from the editor — it only exercises on a real call with a real voice. Test every language before going live.
For each language:
- Call and open in that language — does it reply in the same language?
- Does it stay in that language for the whole call?
- Switch language mid-call — does it follow?
- Say something it must get exactly right — a name, a date, a phone number
- Does it interrupt you when you pause mid-sentence?
- Does the voice sound acceptable to a native speaker?
You cannot judge detection quality in a language you do not speak. A fluent speaker will catch mispronunciations and awkward phrasing in thirty seconds that a transcript will never show you.
Troubleshooting
It replies in English no matter what the caller speaks
Most often the instructions, not the detection. Confirm the language block is present and near the top of Instructions. Then confirm Transcription Language is multi or auto and not a single language code — a hard-set en tells the assistant to expect English and it will try to read other languages as English.
It will not save my language setting
The model and the language have to match. multi only works with English-recommended; language codes outside its ten only work with multilingual-recommended. Check the pairing table above.
Detection is right but the voice sounds wrong
That is the voice, not the detection. See Choosing a voice. Preview your greeting in each language and pick a different voice if needed.
It talks over callers since I switched to Multilingual-recommended
Expected. Multilingual-recommended has no fast end-of-turn detection. Increase Wait Seconds under Speaking Plan, or move back to English-recommended if all your languages are among its ten.
Next Steps
- Voice & Language Settings — voices, timing, and single-language setup
- Writing Instructions — how prompt priority works
- Testing Your Assistant — full pre-launch checklist
Need Help?
Not sure whether your languages fit in one assistant? Contact support and tell us which languages you need and which one most of your callers use. We will tell you which setup fits and can configure it for you.