Deepgram Nova-3
Audio modelCard summarizedOur most advanced speech-to-text model to date
● Re-checked 12 Oct 2026 · 56 facts · 0 corrections · 2 unconfirmedSource: Nova-3 launch post (Deepgram publishes no model card)
Where to call it
- Deepgram API
nova-3
Good for
- Meetings, event captioning, multi-speaker, multilingual, noisy or far-field audio, batch or streaming
- Real-time transcription of code-switched speech in 10 languages
- Keyterm prompting for up to 100 key terms without retraining
Overview
Deepgram · 01Nova, front
Nova-3
Our most advanced speech-to-text model to date
Typical price
$0.0043 / min
Pre-recorded, monolingual
- Max length
- —
- Languages
- Many monolingual models plus multi (10 languages with code-switching)
- Modes
- speech-to-text, streaming
- Released
- 12 Feb 2025
Nova-3 is Deepgram's general-purpose speech-to-text model for batch and streaming audio. The launch post claims a median WER of 6.84% streaming and 5.26% batch on Deepgram's own 2,703-file test set, real-time code-switching across 10 languages, and keyterm prompting without retraining. Deepgram's docs recommend it for meetings, captioning, multi-speaker, multilingual, noisy or far-field audio; it has no turn detection. launch post; models overview
nova-3- Typical price
- $0.0043 / min
- Languages
- Many monolingual models plus multi (10 languages with code-switching)
- Weights
- Closed
- Released
- 12 Feb 2025
- speech-to-text
- streaming
- multilingual code-switching
- keyterm prompting
- diarization
- redaction
- entity detection