Deepgram

Deepgram is a voice-AI company that sells speech-to-text, text-to-speech and voice-agent APIs to developers. It was founded in 2015 by Scott Stephenson, its CEO, and a fellow physicist: they had applied machine learning to waveforms from a dark-matter detector before turning to deep learning for audio at the University of Michigan. On January 13, 2026, Deepgram raised a $130 million Series C at a $1.3 billion valuation, led by AVP, which brought its total funding to more than $215 million. It said more than 1,300 organizations build on its APIs, and Stephenson told TechCrunch the company had been cash-flow positive the previous year. With the round it bought OfOne, a Y Combinator startup that takes drive-through orders by voice.

Products (October 2026)

Deepgram's speed claims center on Flux, which it says makes end-of-turn decisions in under 400 milliseconds, so an agent can reply quickly without cutting people off. That's a vendor figure. At list prices on October 20, 2026, Nova-3 cost $0.0043 a minute for pre-recorded audio (about $0.26 an hour) and $0.0048 a minute streaming, and Flux English $0.0065 a minute. Diarization, redaction and keyterms are priced as add-ons.

Deployment

Deepgram offers more ways to keep audio in-house than most speech APIs. Enterprise customers can self-host its models in Docker or Kubernetes, or run them through Amazon SageMaker. It runs regional endpoints in the EU and Australia, and an India endpoint went generally available on September 15, 2026. Its October 1, 2026 self-hosted release removed Whisper support and added watermarking of Flux TTS output. That flexibility suits contact centers, healthcare and public-sector buyers with data-sovereignty rules.

How It Compares

Deepgram's natural rival is AssemblyAI. At pay-as-you-go list prices, AssemblyAI is cheaper for batch transcription and stronger on multi-speaker, code-switched recordings. Deepgram is cheaper for streaming, has its own text-to-speech and offers more deployment options, which makes it a common base for conversational AI and phone agents. ElevenLabs and Cartesia compete on the voice-output side. Open models such as Whisper and NVIDIA's Parakeet compete for teams willing to run recognition themselves. Every vendor publishes benchmarks it wins, so a blind test on a few hours of your own audio is the only reliable guide. See Deepgram vs AssemblyAI for the detailed comparison.