Nova-3 Multilingual Expands to 55 Languages and Adds Language Hinting
By Steven Van ·
Requests using model=nova-3&language=multi get the new model automatically, and an optional language_hint parameter steers detection toward expected languages.
Deepgram's Nova-3 multilingual model now covers 55 languages. Requests that use model=nova-3&language=multi are served by the new model, per the documentation, on the same /v1/listen endpoints for pre-recorded and streaming audio. Deepgram says no request changes are required.
Language hinting
A new optional language_hint parameter biases detection toward the languages you expect. Repeat it once per language, for example language_hint=en&language_hint=es. Hints bias detection but do not restrict it, so the model still transcribes other supported languages it hears.
- It only applies with language=multi and is ignored otherwise.
- Nova-2 does not support it.
- A hint that is not a supported code fails the request with HTTP 400 (INVALID_QUERY_PARAMETER).
Response changes
Every Nova-3 multilingual response now includes languages_hinted. It holds the normalized hints you sent (en-GB is returned as en), or an empty array when no hint was sent. In streaming, it appears on every interim and final result.
Code-switching and per-word language tags work as before. When the model assigns languages, they come back in languages on each alternative and in language on each word. On audio that switches languages, either field can be absent from a result. For code-switching in streaming, the documentation recommends endpointing=100.
Keeping the previous model
To stay on the earlier 10-language Nova-3 multilingual model, pin it with the version parameter: 2026-01-27.9107 for batch, 2026-01-27.9249 for streaming.