Scott Stephenson

Scott Stephenson is the co-founder and chief executive of Deepgram, the voice-AI company behind the Nova speech-recognition models, Aura speech synthesis, and tools for building conversational agents. A former dark-matter physicist, he develops the infrastructure that makes spoken interactions with AI fast, responsive, and commercially practical.

Stephenson earned a PhD in particle physics at the University of Michigan, where his research included building an underground laboratory to detect dark matter. He left a postdoctoral research position and founded Deepgram in 2015 with fellow Michigan physicists Adam Sypniewski and Noah Shutty. An early motivation was finding useful information within extensive personal audio recordings; the company subsequently joined Y Combinator’s Winter 2016 batch.

Deepgram experimented with searchable audio and video before concentrating on speech models and APIs for other businesses. Call-center analytics supplied an early commercial foothold, and Stephenson initially handled sales himself while developing the company’s enterprise strategy. His account of Deepgram’s early product decisions describes the transition from experimental applications to voice infrastructure.

The company expanded from speech recognition into synthetic speech and complete conversational systems. Its products include Nova-3, Aura-2, Flux, and the Voice Agent API, covering transcription, speech generation, conversational turn-taking, and agent orchestration.

• Context should travel through the entire conversation. Transcripts alone lose tone, emotion, background sound, pacing, and previous exchanges. Stephenson advocates contextual voice AI that passes these signals—and potentially documents, images, or embeddings—between recognition, reasoning, and speech generation.

• Modular voice systems give businesses control. Separating speech recognition, language models, and synthetic speech lets organizations customize safeguards, acoustic adaptation, model size, and vocal delivery. In his AI Engineer World’s Fair appearance, Stephenson highlighted an independently built Daily implementation combining Deepgram speech services with Llama.

• Latency and cost decide whether voice agents scale. Humanlike turn-taking requires rapid processing, while enterprise adoption depends on matching each component’s computational cost to its actual task. Stephenson treats those constraints as core requirements of real-time conversational infrastructure.

In January 2026, Deepgram announced a $130 million Series C financing at a $1.3 billion valuation and acquired restaurant voice-automation company OfOne, extending its reach into enterprise customer-service applications.