Deepgram expands Nova-3 multilingual transcription to 55 languages

Deepgram expanded multilingual language coverage for Nova-3 transcription.

**Dek:** The October 23 update keeps the existing streaming and prerecorded endpoints and adds optional language hints. Applications still need to handle mixed-language results where some language fields are absent.

Deepgram expanded its Nova-3 multilingual speech recognition model to 55 languages on October 23. The existing Nova-3 multilingual model selector now serves the updated checkpoint for prerecorded audio and streaming audio on the same Listen endpoints. It continues to support speech that switches languages mid-conversation. Deepgram’s changelog also documents an optional request hint and a new response field that reports how the service normalized the hint. [Deepgram October 23 changelog](https://developers.deepgram.com/changelog/2026/10/23)

Language hints give the recognizer context

Applications can send one or more expected language codes using the language hint parameter. The field biases detection toward those choices. Deepgram’s example uses English and Spanish for a multilingual request. The response reports normalized hints in the languages_hinted field, such as normalizing a regional English code to English, and returns an empty list when no hints were provided. An application can record that field when it needs to diagnose whether the intended configuration reached the service. [Deepgram October 23 changelog](https://developers.deepgram.com/changelog/2026/10/23)

The language chosen by the recognizer is a different kind of information. Where assigned, Deepgram can include languages on an alternative and a language label on a word. Its changelog says either field can be absent in results from audio that switches languages. Clients consuming those fields should therefore handle absence independently of whether the text transcript is usable. An analytics pipeline that routes a call by language should also avoid treating a hint from the application as proof of what the caller spoke. [Deepgram October 23 changelog](https://developers.deepgram.com/changelog/2026/10/23)

Test the speech patterns the application actually receives

For a call center, a useful evaluation pairs a clean bilingual recording with a real call pattern, such as a caller switching languages while giving an address or reading a product code. Compare hinted and unhinted requests on the same audio and inspect both the transcript and the language fields. Add cases with a short utterance or a dominant language followed by a brief switch, since a total language count does not establish accuracy for a particular pair.

The update also documents a version selector for teams that need to compare behavior with the previous checkpoint. That makes it possible to measure changes with the same input and request settings before changing how an application relies on language labels. The new language coverage is most useful when an operator can see where the updated recognizer improves actual transcripts and where the pipeline still needs a fallback. [Deepgram October 23 changelog](https://developers.deepgram.com/changelog/2026/10/23)