Multilingual Calls and Language Detection

Set a default language, give the agent a voice per language, and let it detect and follow a caller who switches.

Multilingual Calls and Language Detection

A voice agent can run a call in one language, offer several and switch when asked, or work out which language the caller is speaking and follow them without being told. All three are configured from the Setup and Voice tabs.

What you'll need

The agents:voice:manage permission.
A saved voice agent — see Create a Voice Agent.

One language (the default)

On the Setup tab, Default Language sets the language the agent starts in and the language speech recognition is tuned for.

Any language your agent's model and voice support can be the Default Language. The list under Default Language shows exactly those: on a standard model, the languages speech recognition understands AND the chosen voice speaks (OpenAI, Cartesia and Gemini voices each cover a different set); on a Gemini Live model, the languages Live supports. Change the voice or the model and the list changes with it.

If you switch to a voice or model that does not support the language already saved, the language is not changed for you: the form says which part does not support it, and saving is refused until you pick another voice, model or language. Additional Languages rows work the same way, each against its own voice.

Being a supported language is not the same as having a voice built for it. What is supported, language by language explains which engine covers what, and when to choose the voice with care.

Leave Multi-Lingual Support off and you are done; the agent speaks one language in one voice.

Several languages, switched on request

flowchart TB
    A["Multi-Lingual Support: on"] --> B["Default Language<br/>+ primary voice"]
    A --> C["Additional Languages<br/>one voice per language"]
    B --> D["Agent starts in the default"]
    C --> E["Caller asks in, or asks for,<br/>another language"]
    E --> F["Agent switches language<br/>and swaps to that language's voice"]
    G["Auto-Detect Language: on"] --> H["Every turn is recognised<br/>language-agnostically"]
    H --> F

Open the Setup tab and turn on Multi-Lingual Support ("Configure different voices for different languages").
Go to the Voice tab. The voice picker's label changes to Primary Voice — that is the voice for your default language — and an Additional Languages card appears beneath it.
Click Add Language, then for each row:
- Language — pick from the same list.
- Voice — click Choose a voice and pick one, as for the agent's own voice (see Voices and Speech). The list opens on This language, so a French row shows French voices first: Cartesia's, and, with Gemini voices enabled, Google library voices with a French or Canadian French accent. Each row shows its locale, gender and engine, with a play button for your Preview Text.
- Use the trash icon to remove the row.
Save Changes.

Turning on multilingual support also gives the agent its language-switching ability: it can move between the languages you configured when a caller asks, or when it hears the caller has changed. Each switch swaps the voice to that language's voice, so the accent follows the language rather than staying in the default's.

What is supported, language by language

A language works on a call when every part of the call can handle it:

On a standard model (Gemini 3.8 Flash, Gemini 3.7 Flash, GPT-6 Luna), speech recognition has to understand it, and the voice's engine has to speak it. With Auto-Detect Language on, recognition uses a multilingual model that covers fewer languages than the single-language one.
On a Gemini Live model, one Google model listens and speaks, so its own list (99 languages) is all that matters.

The voice engines cover different languages:

Voice	Languages it speaks	Accent
Gemini voices (beta, where enabled)	About 70 with Fast quality, about 80 with Expressive (Finnish, Swedish, Thai and Lithuanian, among others, need Expressive)	Every voice speaks every one of its languages; Google's library adds voices by regional accent
Cartesia voices	44, including Hindi, Punjabi, Arabic, Chinese, Japanese, Korean, Urdu and many European languages	Each voice is built for one language; in another language it keeps its own accent
OpenAI voices (Alloy, Ash, Ballad, Coral, Echo, Fable, Onyx, Nova, Sage, Shimmer)	About 57 (not Punjabi, for example)	Language-agnostic: no accent carried over

You do not need to look any of this up: Default Language and each Additional Languages row list only the languages your agent's model and that row's voice can handle, and the form says which part is missing when a saved language no longer fits. Speech recognition is the part that most often narrows the list on a standard model; a Gemini Live model is the widest.

Voices built for a language exist today in English (16, including US and UK, plus 10 emotive), French (8), Hindi (4), Spanish (2) and Punjabi (2).

A language without a voice built for it: pick the voice deliberately

For German, Italian, Portuguese and every other language without a Cartesia voice built for it, the agent speaks through whichever voice you selected. Two sensible choices:

An OpenAI or Gemini voice: these are language-agnostic, so they speak the language without carrying an English accent across. This is usually the better pick, and these are the voices that also take a Speaking style.
A Cartesia voice built for another language: it speaks the language in that voice's own accent. Preview it before you commit: type a sentence in the target language into Preview Text and listen.

Either way, preview in the actual language rather than the default English phrase. That is the only way to hear what your callers will hear.

Detect the language automatically

On the Setup tab, Auto-Detect Language ("Automatically detect the user's language each turn and switch voices on the fly") changes how speech recognition runs: instead of listening for one specific language, the agent recognises each turn language-agnostically and switches to the caller's language as it hears it.

Use it when:

Callers arrive on one number from several language communities, and asking them to choose is friction you would rather not add.
A caller code-switches mid-sentence — common in bilingual households and worth handling gracefully.

Leave it off when:

Everyone calls in one language. A single-language model is the more accurate choice, and there is nothing to gain from generality.
Your callers use heavy domain vocabulary, brand names or part numbers in one language. Recognition tuned to that one language handles them better.

Auto-detect pairs with Multi-Lingual Support: detection decides what the caller said, and the configured voices decide how the agent answers. Turn on detection without configuring an additional voice for a language and the agent may answer in a language it has no dedicated voice for, in the primary voice.

Writing instructions for a multilingual agent

Write them in one language — English is fine. The model works from them and speaks whatever the call is in; you do not need a translated copy per language.
Say what should not be translated: product names, legal wording, a quoted price format, the business name.
Say what to do when the caller's language is not configured. Offer a transfer, or continue in the default language and say so — either is better than silently answering in a language the caller did not use.
Keep the greeting language-neutral, or in the default. The greeting is spoken before the agent has heard anything, so detection has not run yet.
On outbound calls, the opener is in the call's language without extra work. When the person who answers says nothing, the agent opens with your Static Greeting, or else with the opener its pre-call plan wrote. Without either, the agent writes its own opening line in the call's language from your instructions. If that turn fails, it speaks the platform's line, translated into the call's language; only when no translation can be had does it fall back to English. Write a Static Greeting only when you want exact words. The call’s detail page shows which opener was used, in its First turn row.

Live-audio models: one voice, a checklist of languages

On Gemini 3.8 Live the per-language voices and Auto-Detect Language above do not apply: one Google voice speaks every language and the model recognises the caller's language itself. Instead, Voice → Languages the agent may speak is a checklist of the languages Google's Live models support (99 of them). The agent answers in the caller's language when it is ticked, otherwise in the first ticked one; with nothing ticked it keeps to the Default Language. An agent that was multilingual before you moved it to a Live model keeps its languages until you tick the list. Portuguese is the one language the list splits by region (Brazil, Portugal).

Troubleshooting

Saving fails with "… is not supported by …". The voice or model you picked cannot run in the Default Language (or in an Additional Languages row). The message names the part: pick a voice whose engine speaks it, a different model, or another language.
The agent speaks the language with an English accent. The voice you selected was built for another language. Switch to an OpenAI or Gemini voice, which is language-agnostic, and preview it with a sentence in the target language.
A language I expected is not in the list. The list holds what your model and voice can both handle. Try a Gemini voice (or its Expressive quality), turn Auto-Detect Language off, or use a Gemini Live model, whose list is the widest.
The Additional Languages card is not there. Multi-Lingual Support is off, on the Setup tab.
The agent switches language but keeps the original accent. No voice is configured for the language it switched to. Add a row for it under Additional Languages.
It answers in the wrong language on the first turn. The greeting and the opening turn run in the default language — detection needs to hear the caller first. Set the default to your most common language.
Recognition got worse after turning on auto-detect. Expected on a single-language line: a language-agnostic model trades some accuracy for coverage. Turn it off if every caller speaks the same language.

Multilingual Calls and Language Detection