Deepgram is a high-accuracy, low-latency speech-to-text API built for real-time and batch transcription at scale.
✓ Pros
- Best-in-class real-time transcription latency (<300ms)
- Nova-2 model delivers top accuracy on noisy audio
- Speaker diarization, smart formatting, and topic detection included
- Generous $200 free credit on signup
✗ Cons
- Multilingual support still narrower than Azure Speech or Google STT
- On-premises deployment only on Enterprise tier
- No built-in meeting recorder—API-only product
| Free tier | ✓ Free tier |
| Pricing model | usage |
| Price (1 hour) | $0.10 USD |
| Features | |
| Languages | en, ja |
| API | ✓ Available Docs ↗ |
| Pricing Plans | Free$0$200 in free credits on signup Pay-as-you-go$0.0043/minNova-2 model, no commitment GrowthFrom $4,000/yrVolume discounts, dedicated support EnterpriseCustomOn-prem, SLA, custom models |
| Platforms | |
| Integrations | Twilio, Vonage, AWS, WebSocket streaming, Node.js / Python SDK |
| Homepage | https://deepgram.com |
AI Commentary
Deepgram's Nova-2 model consistently ranks at or near the top of independent STT benchmarks for accuracy and latency on English audio. Its WebSocket-based real-time streaming is a preferred choice for live captioning, call center analytics, and voice-first application developers. The platform's developer experience—comprehensive SDKs, good documentation, and a generous free tier—has built a strong community. Multilingual breadth remains a gap versus Azure Speech.