Multilingual Setup Guide

Each language is identified by a BCP-47 code. Set this code in your dashboard or config to lock the agent to that language. The special value auto lets the model detect and match the caller's language automatically.

There are three ways to set the language. The dashboard is the simplest for most users. All methods produce the same result - the agent speaks and listens in that language.

# Example: set Hebrew via environment
NINJA_TALK_LANGUAGE=he-IL
NINJA_TALK_VOICE=Kore

All 30 available voices work in all supported languages. However, some voices sound more natural in certain languages. The table above lists recommended pairings based on production testing.

For gendered languages (Hebrew, Arabic, Russian, Polish, Hindi), the voice's gender determines the bot's grammatical forms. A female voice like Kore will use feminine first-person conjugation. A male voice like Charon will use masculine forms. This happens automatically - no extra configuration needed.

For best results, write the entire playbook (script, greeting, and tool descriptions) in the target language. The model produces more natural output when its instructions match the expected output language.

If you set the language to he-IL but write the playbook in English, the bot will still respond in Hebrew - but phrasing may sound translated rather than native. Writing in the target language from the start eliminates this.

// Example Hebrew greeting in dashboard:
"shalom, ani ha-sochenet ha-virtualit shel [Company].
eich ani yechola laazor lecha hayom?"

// The greeting field accepts text in any language.
// Write it exactly as you want the bot to speak it.

When a caller switches languages mid-call (common in multilingual markets like Israel, India, or Quebec), the bot's behavior depends on your language setting:

• auto - the bot follows the caller and switches to match. Best for markets where code-switching is common.
• Fixed language (e.g. he-IL) - the bot stays in Hebrew regardless. Even if the caller mixes in English loan-words, the bot responds in Hebrew. The language rules explicitly instruct: "Never code-switch."

For businesses that must communicate in a specific language for legal or brand reasons, set a fixed language. For general customer service where callers may speak different languages, use auto.

When placing outbound calls via the Voice API, you can effectively override the language on a per-call basis by writing the task goal, script name, and facts in the target language. The bot will speak in whatever language it receives its instructions in.

POST /v1/voice/calls
{
  "from": "+16473705600",
  "to": "+972501234567",
  "triggeredBy": "billing-system",
  "task": {
    "goal": "להזכיר ללקוח על חיוב פתוח",
    "script": "payment-reminder-he",
    "facts": {
      "שם_לקוח": "דוד כהן",
      "סכום": "250 שקלים"
    }
  }
}

The deployment-level language setting still applies as the default. Per-call instructions in a different language override it for that specific call.

Hebrew (he-IL) and Arabic (ar-EG) are right-to-left languages. On the voice side, RTL has no impact - speech is speech. However, there are considerations for the text portions of your setup:

• Playbook editing - the dashboard text fields accept RTL text natively. Write Hebrew or Arabic directly.
• Caller gender detection - both Hebrew and Arabic have rich grammatical gender. The bot auto-detects the caller's gender from their first 1-2 utterances and locks verb/adjective agreement for the rest of the call.
• Bot self-reference - in Hebrew and Arabic, the bot's own gender forms are determined by the selected voice. Female voices (Kore, Aoede, Sulafat) use feminine self-reference. Male voices (Charon, Puck, Orus) use masculine.
• Numbers and dates - the bot handles mixed LTR/RTL content (phone numbers, dates, amounts) automatically within speech.