Your users speak 60+ languages. So do we.
Transcribe, understand, and speak the languages your customers use, with production accuracy and low latency.
From hello to hola to ciao, and more.
Production speech-to-text in 60+ languages, with text-to-speech in 45+ of those languages and 55+ regional variants.
Coverage that holds up in production
- 60+Languages and markets
- 45+Languages with real-time code-switching
- 4In-region endpoints, with more in progress
Multilingual or monolingual, your choice
Use monolingual when the language is fixed, and multilingual when callers may switch as they speak.
Full conversations in 45+ languages
Deepgram Flux STT Multi and Deepgram Nova-3 Multi cover English, Spanish, French, German, Japanese, Italian, Dutch, Portuguese, Hindi, and Russian.
Switch mid-sentence
Callers change language mid-call and one model carries the whole conversation, holding its state straight through the switch.
Easily go global
One agent serves every market, which matters when the caller's language is unknown before they speak.
A model per language
Nova-3 runs a dedicated model for more than 70 languages and markets, tuned specifically for that one language.
Highest accuracy in noise
Knowing the language before the call starts lets the model optimize around it, which performs well in noisy and domain-specific audio.
Declared, auditable language
Regulated workloads that have to state which language was processed get an explicit declaration, rather than inference.
Models for multilingual voice AI
Choose the model that fits your workload, from conversation-aware speech recognition and production transcription to natural, multilingual speech generation.
Flux STT
Conversation-aware speech recognition for multilingual voice agents, with turn-taking built into the model.
Nova-3
Production transcription for global applications, across dedicated and multilingual models.
Flux TTS
Conversation-native speech synthesis for multilingual voice agents, with expressive delivery across an expanded language set.
Aura-2
Flexible multilingual speech synthesis for dependable production voice experiences.
FAQs
More than 60 languages and markets, across speech-to-text and text-to-speech. Coverage varies by model and use case, so see our speech-to-text language docs and text-to-speech language docs for current availability.
A multilingual model handles multiple languages in one session, so callers can switch language mid-conversation without a handoff. A monolingual model is dedicated to one language you declare upfront and tuned for it, which gives the highest accuracy in noisy or domain-specific audio.
Yes. Flux STT Multi handles code-switching natively, with optional language hints to bias detection. Nova-3 Multi supports code-switching through the language=multi setting.
For the Flux models, voice agents can speak and understand English, Spanish, French, German, Dutch, Italian, Japanese, Portuguese, Hindi, Russian, and Mandarin Chinese. See the speech-to-text and text-to-speech language pages for the full list and model-specific availability.
Yes, across 35 languages for pre-recorded audio. For real-time streaming where the language is unknown or changes, the multilingual models are the better fit, since they identify and transcribe language shifts inside the session rather than deciding once at the start.
Yes. Deepgram offers regional endpoints in the U.S. (Global), EU, Australia, and India. Use your existing API keys and SDKs, and switch the base URL.