Learn the best way to support a multi-lingual voice agent through Deepgram.

Select a multilingual speech-to-text (STT) model.

In order to have a fully multi-lingual agent, you need to select a multi-lingual speech-to-text model and specify the appropriate language. Today, you should specify the following parameters in your Settings.

agent.listen.provider.model : nova-3
agent.listen.provider.language : multi

Select a multilingual text-to-speech (TTS) provider.

OpenAI, Eleven Labs, and Cartesia have multi-lingual language models. Select one of those providers and use agent.speak.provider.language as multi. For Eleven Labs, this parameter aligns to their language_code.

Prompt Recommendations

Prompt design can help adjust the agent’s behavior depending on your expected use case. Results may vary depending on the LLM provider you use.

Use Case	Recommendation
Mirroring conversation that switches back and forth between two languages (e.g. English > Spanish > English).	To more reliably mirror the end user’s language, provid eexplicit per-turn language instructions (e.g. “match the language of each user message independently”).
Enforce an agent to speak a single language (e.g. English) , even if a user speaks a different language (e.g. Spanish).	Use a strict English-only instruction in the prompt.
Provide bilingual support where relevant (e.g. English input in English-speaking agents, Spanish input results in natural bilingual Spanish and English).	Specify conditional language mixing in the prompt, (e.g. “Respond in English unless the user speaks Spanish; if Spanish, you may mix Spanish and English naturally”).