Scott Stephenson is Deepgram’s CEO and co-founder. He moved from particle-physics research into building speech recognition and voice AI infrastructure.[1][2]
Start with How to test automatic speech recognition providers →
Background
Stephenson earned a doctorate in particle physics at the University of Michigan. His company biography describes research on detecting dark matter in an underground laboratory and says he left a postdoctoral research position to found Deepgram. The connection to voice technology was signal analysis: finding useful patterns in large amounts of recorded information.[2][3]
Career
Deepgram’s founder account on Y Combinator dates the company to 2015 and names Stephenson, Adam Sypniewski and Noah Shutty as its founders. It describes an early practical problem: wearable recording devices produced hundreds of hours of conversations and notes that were difficult to search. Stephenson built a deep-learning model for finding information in voice data. The company’s account traces its subsequent move toward enterprise speech recognition, where customers need software to turn recordings into usable text.[3]
Contributions
Deepgram now offers speech-to-text, text-to-speech, audio intelligence and a voice-agent API. These components address different parts of a conversation: recognizing a person’s words, extracting information from audio, and producing a spoken response. Stephenson’s public contribution includes leading the company and explaining its product direction. His bylined introduction of Nova in 2023 describes a speech-recognition release and its evaluation on varied audio. The performance comparisons are Deepgram’s own measurements; they are useful for understanding its engineering priorities without treating a launch benchmark as a universal ranking.[1][5]
Ideas
Stephenson’s guide to choosing speech-recognition providers begins with the customer’s requirements. It asks readers to examine batch and real-time speed, support for multiple audio channels, and separation of different speakers. That is a useful starting point for a builder: a transcript can contain mostly correct words and still be unsuitable if it arrives too late or cannot identify who said what. His 2026 essay extends this concern to an “audio Turing test,” a phrase he uses for a spoken interaction that feels indistinguishable from conversation with another person. He presents this as a direction for development, not a completed scientific result.[4][6]
Tools, writing & talks
Writing, research & talks
Start here
A practical checklist for evaluating transcription services against an actual workflow. Start here to distinguish word accuracy from response time, channel support and speaker separation before choosing a voice stack.
A concrete checklist for understanding what makes speech recognition useful, without requiring machine-learning research experience.
Author[4]- Introducing Novadeepgram.com
Stephenson’s launch explanation for Deepgram’s 2023 speech-to-text model. It describes the company’s accuracy, speed and cost comparisons and its approach to training across audio domains. Read the benchmark claims as vendor-reported results.
Author[5] - The next chapter of the voice AI economydeepgram.com
A 2026 essay linking Deepgram’s company direction to natural spoken interaction. Its discussion of active listening and the audio Turing test makes the ambition understandable while leaving the distinction between aspiration and demonstrated capability intact.
Author[6]