> ## Documentation Index
> Fetch the complete documentation index at: https://docs.anyreach.ai/llms.txt
> Use this file to discover all available pages before exploring further.
# Multi-language
> Configure the languages an agent understands and speaks.
The agent's `language` block declares which languages it can understand and speak. It accepts a list of language codes, and the platform validates that list against the rules below and against the language support of the speech provider you select. Speech-to-text, text-to-speech, and turn detection all read from this configuration.
## The language block
`language` is an optional object on the agent. When set, it requires a non-empty `languages` array.
| Field | Type | Default | Description |
| ----------- | ---------- | ------- | ------------------------------------------------------- |
| `languages` | `string[]` | none | List of language codes. At least one entry is required. |
```json theme={null}
{
"language": {
"languages": ["en-US"]
}
}
```
A single-language agent may use either a base code (`en`) or a locale-specific code (`en-US`). A multi-language agent must use base codes only.
## Base codes vs locale codes
A base code names a language (`en`, `es`, `fr`). A locale code adds a region with a hyphen (`en-US`, `pt-BR`, `es-419`).
The rule is enforced by the agent language validator: if `languages` contains more than one entry, no entry may contain a hyphen. A locale-specific code is only valid when exactly one language is selected.
| Selection | Allowed | Example |
| ------------------ | ------------------- | ----------------------- |
| Single language | Base or locale code | `["en"]` or `["en-US"]` |
| Multiple languages | Base codes only | `["en", "es", "fr"]` |
Mixing a locale code into a multi-language list is rejected. `["en-US", "es"]` fails validation with `Locale-specific code 'en-US' not allowed with multiple languages. Use base language code instead.` Use `["en", "es"]`.
## Per-provider language support
Each speech provider enforces its own supported-language set. Pick languages that the provider you configure on the agent's stack actually supports. See [/agents/voice-and-model-config](/agents/voice-and-model-config) for how providers are assigned.
| Provider | Role | Supported languages |
| ------------------ | ---- | ----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- |
| Deepgram | STT | 48 base codes (`ar`, `en`, `es`, `fr`, `de`, `hi`, `ja`, `zh-HK`, and more). Nova-3 also accepts locale variants such as `en-US`, `pt-BR`, `es-419`, and the special `multi` value. |
| Gladia (Solaria-1) | STT | 101 base codes. Each entry in `languages` is validated against the supported set; an unsupported code raises `Unsupported Gladia language code`. |
| Cartesia (Sonic-3) | TTS | 42 languages. Defaults to `en`. |
The Deepgram and Gladia language lists are validated server-side. Submitting a code outside a provider's supported set is rejected at save time, not silently dropped.
### Deepgram multi-language mode
Deepgram Nova-3 supports a `multi` language value that lets a single STT stream recognize multiple languages. In this mode, key terms are disallowed: providing `keyterms` together with `language: "multi"` raises `keyterms can only be provided when language is not specified as 'multi'`. Set `keyterms` only when the language is a specific code, not `multi`.
### Gladia code-switching
Gladia Solaria-1 accepts a `languages` list and a `code_switching` flag (default `true`) that lets recognition switch between the listed languages mid-utterance. An empty `languages` list enables auto-detection.
## The anyreach meta provider
The `anyreach` STT provider is a meta provider: it does not bind to one engine. At runtime it resolves to Deepgram or Gladia based on the agent's configured language, so you do not have to choose the underlying STT engine per language yourself.
Its `default` model takes an optional `vocabulary` list of domain-specific terms. That vocabulary maps to Deepgram `keyterms` or Gladia `custom_vocabulary` depending on which engine is resolved.
```json theme={null}
{
"provider": "anyreach",
"model": {
"name": "default",
"parameters": {
"vocabulary": ["AnyReach", "PostgREST", "LiveKit"]
}
}
}
```
## Multilingual turn detection
Turn detection decides when the caller has finished speaking. The `anyreach` turn-detection provider offers a `multilingual` model that adapts to the agent's language rather than assuming English, so end-of-turn timing works across the languages the agent supports. LiveKit's turn detection exposes both an `english` and a `multilingual` model for the same reason.
Pair multilingual turn detection with a multi-language agent and a multilingual-capable STT provider so every stage of the pipeline handles the same set of languages.
## Putting it together
One language lets you use a locale code (`en-US`) for region-specific recognition and speech. Multiple languages require base codes (`en`, `es`).
Confirm each code is in the supported set for the STT and TTS providers on the agent's stack. Use the `anyreach` STT provider to let the platform resolve the engine for you.
For multi-language agents, use a multilingual turn-detection model so end-of-turn timing adapts to the spoken language.
Add `keyterms` (Deepgram) or `vocabulary` (anyreach) only when targeting a specific language. Deepgram `multi` mode rejects key terms.
Assign STT, TTS, and turn-detection providers to the agent stack.