> For clean Markdown of any page, append .md to the page URL.
> For a complete documentation index, see https://docs.rydhm.ai/llms.txt.
> For AI client integration (Claude Code, Cursor, etc.), connect to the MCP server at https://docs.rydhm.ai/_mcp/server.

# Languages

An expert can talk in more than one language. The **Languages** setting tells Rydhm which ones your callers really speak. It does two things at once:

1. It tells the **transcriber** which languages to expect, so it hears them accurately.
2. It tells the **model** that it may answer in those languages, so it replies in the caller's language.

## Choose the languages

#### Open the Voice tab

Open the expert, then **Voice**. **Languages** is in the **Voice** section, under the voice.

#### Tick the languages

Choose any of these ten:

|         |         |         |            |          |
| ------- | ------- | ------- | ---------- | -------- |
| English | Hindi   | Spanish | Portuguese | French   |
| German  | Italian | Dutch   | Russian    | Japanese |

An expert always has at least one language. One that was never set behaves as English.

#### Check the transcriber

Rydhm moves the **Transcriber** (under **Advanced**) with your languages, so the two can't disagree:

* Two or more languages switch *Deepgram Flux — English* to *Deepgram Flux — multilingual (code-switching)*, and *Deepgram Nova-3 — English* to *Deepgram Nova-3 — multilingual*.
* Going back to one language switches back to the English model, which takes turns faster.

#### Save and test

Choose **Save changes**, then talk to the expert in each language from the **Test** panel.

> **Tip**
>
> **Pick only what is really spoken.** A language you add is a hint, not a filter: each extra language costs a little accuracy on the others.

## Languages outside the list

The ten languages above are the ones Rydhm's multilingual transcriber hears in both calls and meetings, so a language you tick works everywhere the expert goes. For other languages, choose **Gemini 3.5 Transcribe Live — any language, auto-detected** as the expert's **Transcriber** (Voice tab → **Advanced**). It detects the language of each thing a caller says, and is the only option that hears Malayalam, Gujarati, Bengali, Tamil, Telugu, Kannada or Marathi. Its turn-taking rhythm is a little different from Flux's.

## How each engine handles languages

| Engine                      | What happens                                                                                                                                                                                       |
| --------------------------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| **Standard** (Balanced)     | The transcriber hears the languages you chose. The model answers in the caller's language, and the voice reads each reply in the language it is written in.                                        |
| **Gemini Live** (Fastest)   | The model switches language by itself once you've allowed it. Its voices speak whatever language the conversation is in.                                                                           |
| **GPT-Live** (Most natural) | The model hears and answers on its own. A few of its voices are made for one language (Hindi, German, French, Japanese, Korean); pick one of those for an expert that mostly speaks that language. |

See [Voice engines](/experts/voice-engines).

## Pick a voice that speaks the language

Every voice in the library has a first language and may speak others. On the **Voices** page, pick a language and turn on **Native speakers** to see only voices whose first language it is. When an expert speaks one language other than English, **Change voice** opens already filtered to it. The sample on the expert's voice card plays in the expert's first language when the voice speaks it, and **Hear the first message** reads the greeting in the language it is written in.

**Tone** is exact only on Cartesia voices in English. In other languages it is ignored or followed loosely. See [Voices](/experts/voices#make-the-voice-sound-right).

## What visitors see

On a [share link](/channels/public-links) and the [website widget](/channels/web-widget), an expert with two or more languages shows a language picker. The visitor's choice is the language the expert opens in, and the expert follows if they switch later. The choice is remembered for that link in the visitor's browser.

On the widget you can set the starting language yourself with the `language` attribute, a two-letter code the expert speaks:

```html
<rydhm-expert expert="<link>" language="hi"></rydhm-expert>
```

The visitor can still change it.

## Gotchas

* **Write the first message in the language most callers expect.** It is spoken as written.
* **Templates and Describe it set languages for you.** Check them after creating an expert.
* **The knowledge search matches English words best.** Knowledge in other languages is still found by meaning. See [Search and sources](/knowledge/search).

## Related

* [Voice engines](/experts/voice-engines)
* [Voices](/experts/voices)
* [Persona and first message](/experts/persona)