From the archive
Deepgram releases Nova-3 for recorded and streaming speech
Nova-3 launches with hosted English transcription and keyterm prompting for specialized vocabulary.
Deepgram released Nova-3 for hosted English speech transcription. At launch, the model supported both recorded audio and real-time streaming through its application programming interface (API).
Vocabulary without retraining
Keyterm prompting gives the transcription service a list of important terms, such as product names or specialist vocabulary. The launch supported up to 100 terms without requiring a separately trained model.
Other announced improvements focused on:
- Recognition in background noise and overlapping speech.
- Number formatting and specialized terminology.
- More precise word timestamps and transcript formatting.
These features are useful for call recordings and live conversations where a generic transcript can miss the words that matter most.
English first
The initial release was the hosted English model. Deepgram said multilingual and self-hosted deployments would follow, so those options were not part of the same launch availability.
The Deepgram profile covers its hosted speech services and pricing model. For a downloadable transcription model, the Whisper profile describes the local-running option.