Curiosity · AI Model
Deepgram Nova-3
Deepgram Nova-3 is Deepgram's third-generation flagship ASR model, purpose-built for real-time streaming transcription. It delivers sub-300 ms end-to-end latency, built-in diarisation, and keyterm prompting that boosts accuracy on domain jargon — the reason it anchors many voice-agent and contact-centre stacks.
Model specs
- Vendor
- Deepgram
- Family
- Nova
- Released
- 2025-02
- Context window
- 1 tokens
- Modalities
- text, audio
- Input price
- n/a
- Output price
- n/a
- Pricing as of
- 2026-04-20
Strengths
- Sub-300 ms real-time latency
- Built-in diarisation and keyterm prompting
- Strong accented-English WER versus Whisper-class models
- Self-serve API plus dedicated contact-centre connectors
Limitations
- Fewer languages than Whisper large-v3 (though expanding)
- Closed API only — no self-hostable weights
- Per-minute pricing can add up on long archives
- Nova-3 tuning favours English and major EU languages
Use cases
- Voice agents and realtime AI phone calls
- Live captioning and conference subtitles
- Contact-centre analytics and QA
- Meeting-notes pipelines with diarisation
Benchmarks
| Benchmark | Score | As of |
|---|---|---|
| English WER (Nova-3) | ≈6.8% | 2025 |
| Real-time latency p95 | <300 ms | 2025 |
Frequently asked questions
What is Deepgram Nova-3?
Nova-3 is Deepgram's third-generation speech-to-text model, designed for streaming real-time transcription. It emphasises low latency, built-in diarisation, and keyterm prompting to boost accuracy on domain-specific vocabulary.
How is Nova-3 different from Whisper?
Whisper is open-weight and great for batch transcription, but was not built for streaming. Nova-3 is closed-API-only but purpose-built for real-time voice agents, contact centres, and live captioning, with sub-300 ms end-to-end latency.
What is keyterm prompting?
Keyterm prompting lets you pass a list of domain-specific terms (product names, medical terminology, people names) with each request. The model biases decoding toward these terms, improving WER on jargon-heavy audio.
How is Deepgram priced?
Deepgram bills per minute of audio processed, with separate rates for streaming versus batch. Enterprise contracts offer volume discounts, HIPAA, and on-prem options.
Sources
- Deepgram — Nova-3 announcement — accessed 2026-04-20
- Deepgram — API docs — accessed 2026-04-20