Languages

The seven languages an agent can speak, which hearing and voice options work for each, and what to test.

What you will achieve

You will pick the right language for an agent and understand how Phynxa sets up hearing, voice and model for it.

Supported languages

Language	Direction	Notes
English	Left to right	The default. All hearing and voice options work.
Arabic	Right to left	Modern Standard Arabic. Regional dialects are understood to varying degrees.
Urdu	Right to left	The newest language here. Test with real callers before relying on it.
Hindi	Left to right	Hindi and English mixed in one call works best with Deepgram Nova 3 Multilingual.
French	Left to right
Spanish	Left to right
Chinese (Mandarin)	Left to right	Simplified characters.

Step-by-step setup

Open the agent and the Assistant tab.
In Language and timezone, choose the Language.
Check the Hearing and Voice cards. Options that cannot handle the language are greyed out with a reason, such as "cannot listen in Urdu".
Update the Opening line and System prompt to the language you chose, and press Save.

Configuration details

Hearing. Deepgram Nova 3 Multilingual does not support Arabic, Urdu or Chinese; use Deepgram Nova 3 (single language) or GPT-4o Mini Transcribe for those.
Voice. Cartesia, ElevenLabs and OpenAI can all speak every language, but OpenAI's voices are tuned for English and will carry an accent. For Urdu, ElevenLabs uses its v4 Turbo model.
Model strength. The model picker rates each model for the chosen language. Small models are weak outside English; the picker suggests a stronger one (for example GPT-4.1) when the choice is poor.
Cost. Languages other than English use more model tokens. The editor shows the factor, for example "1.6× English tokens" for Hindi.
Turn-taking. For Urdu the agent waits for a short pause to decide the caller has finished, because the smarter end-of-turn model was not trained on it.
Reply language. The agent always replies in its Language, English included, even when a caller speaks another language or a knowledge document is written in one. For an agent that should follow the caller's language, choose Deepgram Nova 3 Multilingual under Hearing.
Timezone. The same card has Timezone. It is the agent's clock: it tells the agent the current time, so "tomorrow at 3" means the right day. It does not tell the agent where your business is. A Karachi timezone on a London company's agent changes the time it uses, never the city it names.

Business details

The Business details card holds the facts the agent states as correct: Business name, Office address and Opening hours. They take precedence over anything in a knowledge base.

Set the Office address if callers ask where you are. Without one, the agent answers location questions only from your knowledge base, and otherwise says it does not have the address, rather than guessing.
Keep these in the card rather than only in the system prompt: they are sent to the agent as verified facts on every call and in the test chat.

Warning. Quality varies by language, model and caller. Always test with a native speaker before publishing, especially in Urdu and Arabic.

Testing

Use Talk and speak the language the way real callers do, including a mixed-language sentence if your callers do that.

Expected result

The agent hears you correctly, answers in the language with a natural voice, and the transcript shows the text in the right script and direction.

Troubleshooting

A voice or hearing option is greyed out. It cannot handle the language; pick another.
The agent mishears. Try Deepgram Nova 3 instead of GPT-4o Mini Transcribe, or the reverse.
It answers in a language you did not choose. Check Language, then Hearing: Deepgram Nova 3 Multilingual makes the agent follow the caller's language. Also check the system prompt does not ask for another language.
It names the wrong city or address. Set Office address under Business details. The timezone is never used as a location.

Something here does not match what you see in the dashboard? Let us know.