Realtime API

Languages

Name the language of the audio, or let the model detect it.

Name the language, or let the model detect it

Automatic detection is the default: leave the language out and the model detects the language being spoken. When you know it, name it, because naming the language is more accurate. For Brazilian Portuguese:

"transcription": {"model": "nemotron-asr-streaming", "language": "pt-BR"}

Set the language before the first audio; it holds for the whole session. With a client secret, set it when you create the secret.

On the Deepgram-compatible API

Pass the same tag as the language query parameter. Leave it out, or set language=multi, and the model detects the language.

Supported languages

Name one of these tags. A tag the model does not take is refused with unsupported_language.