Overview

Deepgram sells speech recognition as infrastructure rather than as a product with a UI. Nova-3 is the general model, priced per minute with splits between monolingual and multilingual and between batch and streaming: $0.0043 and $0.0048 per minute respectively on pay-as-you-go for English. Flux is the newer and more interesting piece, a streaming model with turn detection built into the transcription rather than bolted on top, so a voice agent stops interrupting people mid-sentence. Aura-2 covers text to speech and a Voice Agent API at $0.056 per minute wraps the whole loop for teams that do not want to assemble it.

For anyone operating across languages the 2026 Asia-Pacific expansion is the change that matters. Nova-3 added Mandarin Simplified, Mandarin Traditional and Cantonese Traditional, with a claimed 65.21% relative word error reduction on Mandarin Simplified over Nova-2 and 24.82% on Cantonese. French and German have been solid for considerably longer. Real-time multilingual with code-switching does exist, but the independent numbers put it last among the three big vendors at mixing languages inside a single sentence, so a French engineer dropping into English technical terms mid-thought will generate errors somebody has to fix.

The right buyer has volume and a deployment constraint. Contact-centre analytics, meeting capture at scale, voice agents where per-minute cost compounds fast or any pipeline where the audio has to stay inside your own network. The wrong buyer wants one vendor for everything, because Aura-2 will not satisfy anyone who has heard Sonic-3.6 or Eleven v3, and low-volume high-accuracy work is better served by paying AssemblyAI or ElevenLabs and not thinking about it again.

Key Features

✓ Nova-3 speech to text with separate batch and streaming endpoints, priced per minute from $0.0043

✓ Real-time multilingual transcription with code-switching, covering Mandarin Simplified, Mandarin Traditional and Cantonese Traditional after the 2026 Asia-Pacific expansion

✓ Flux, a streaming conversational STT model with end-of-turn detection built in, so an agent stops guessing when the human finished speaking

✓ Keyterm prompting for product names, part numbers and jargon without training a custom model

✓ Aura-2 text to speech and a Voice Agent API at $0.056 per minute that wraps the full STT to LLM to TTS loop

✓ Self-hosted and VPC deployment, with documented patterns alongside AWS, NVIDIA and Fortanix

My Take

The accuracy lead is gone. On the 2026 code-switching benchmark Nova-3 Multilingual posts 12.22% WER against AssemblyAI Universal-3.5 Pro at 7.69%, and diarization is worse still at 37.92%. What Deepgram kept is the pair of things that usually decide a production deployment: unit price and where the model is allowed to run. Streaming at roughly $0.29 an hour undercuts AssemblyAI by about a third, and self-hosted is a documented deployment path rather than an enterprise promise, which is the entire argument for anyone transcribing audio that cannot leave a plant network or a hospital VPC.