> ## Documentation Index
> Fetch the complete documentation index at: https://docs.qall.io/llms.txt
> Use this file to discover all available pages before exploring further.

# Multilingual calls

> Taking calls in more than one language.

If your callers don't all speak the same language, there are three things to set
and one trap to avoid.

## 1. Transcription

Set the assistant's **language** to the one you expect most.

For a line that genuinely mixes languages, use **Soniox** for speech to text and
fill in **language hints** with the languages you expect. It can follow a caller
who switches mid-sentence, which a single-language setting can't.

Keep the hint list short. Listing eight languages makes every one of them
slightly harder to recognise.

## 2. The voice

Pick a voice that natively speaks the language your callers use most. A voice can
usually pronounce another language, but the accent gives it away — and on a phone
line, accent is most of what someone judges.

If one line serves two languages with equal weight, consider two assistants on
two numbers instead of one assistant doing both. It's usually better for the
caller and easier to tune.

## 3. The prompt

The assistant replies in the language the caller uses. You don't need to tell it
to — but you do need to make sure nothing in your setup forces one language.

Two things commonly do:

| | |
| - | - |
| **First message** | Written in one language. It's spoken verbatim, so a German caller hears your English greeting. |
| **Idle message** | Same. Leave it empty and the assistant checks in in whatever language it's been speaking. |

The same applies to any **Say** step in a [playbook](/playbooks/structured) —
those are spoken word for word, in the language you wrote them.

## The trap: consent

If you ask for [recording consent](/guides/recording-and-consent) and leave the
consent prompt empty, Qall uses stock wording — and it only has stock wording in
English and German. For any other language, **the call isn't recorded at all**.

Write your own consent prompt for every language you take calls in.

## Getting names and places right

Transcription struggles most with proper nouns, and more so across languages. Add
them to **keyterms** in the speech settings — your hotel names, street names,
product names. It's five minutes and it fixes the thing prompts can't.

## Checking it works

Call the assistant yourself in each language. The [Accuracy](/calls) figure on a
call tells you how confident the transcription was — if it's low for one language
and fine for another, that's where to work.


This documentation is built and hosted on [Mintlify](https://mintlify.com), a developer documentation platform.