Overview
Deepgram is speech AI infrastructure rather than a consumer transcription app. It sells APIs: Nova-3 for speech-to-text, Aura-2 for text-to-speech, and a Voice Agent API that bundles listening, LLM orchestration and speaking behind one WebSocket connection. That packaging matters because most teams building phone agents currently stitch three vendors together and inherit three sources of latency. Nova-3 covers 45+ languages with speaker diarization, smart formatting, keyterm prompting and automatic language detection, and Deepgram bills per second rather than rounding audio up to the nearest 15 seconds - a detail that quietly saves 10-20% on short clips. In 2026 the company added Flux, a conversational STT model with built-in turn detection and interruption handling aimed specifically at real-time agents. Deployment options include cloud, VPC and fully self-hosted, which is why Deepgram turns up in healthcare and call-center stacks where audio cannot leave the perimeter. The trade-off is that it gives non-developers nothing out of the box: no meeting-notes UI, no collaborative editor, no shared workspace. You get endpoints, SDKs, a console and generous free credit. If you want a finished product, buy a transcription app; if you are building one, this is the layer underneath it.
Key Features
- Nova-3 speech-to-text across 45+ languages with diarization and smart formatting
- Flux conversational STT with native turn detection for real-time voice agents
- Aura-2 text-to-speech billed per 1,000 characters
- Unified Voice Agent API combining STT, LLM orchestration and TTS over one socket
- True per-second billing with no rounding up of short audio
- Self-hosted and VPC deployment for regulated audio workloads
- Audio Intelligence add-ons for summarization, topic detection, sentiment and intent
Pricing
| Plan | Price | For |
|---|---|---|
| Pay As You Go | $200 free credit, then from $0.0077/min (Nova-3 mono) | Developers and prototypes |
| Growth | $4,000+/year prepaid, up to 20% off | Production apps at scale |
| Voice Agent API | $0.050-$0.163/min depending on BYO components | Real-time phone and web agents |
| Enterprise | Custom | High volume, self-hosted, SLAs |
Comparison
Compared to Whisper, Deepgram trades open weights for managed real-time streaming, diarization and an SLA - Whisper is free to self-host but you own the GPU bill and all the latency engineering. Compared to ElevenLabs, the two only overlap on text-to-speech: ElevenLabs wins on voice cloning and expressive delivery, Deepgram wins on cost per character and on keeping transcription in the same account. Vapi sits one layer higher and can call Deepgram as its transcription provider, so the real decision is buy-the-orchestrator versus build-on-the-API.