DEEPGRAM

Scott Stephenson was a particle physicist. He worked on experiments buried deep underground, hunting for dark matter and neutrinos.

That means sifting tiny real signals out of enormous piles of noise. In 2015 he teamed up with fellow physicists Adam Sypniewski and Noah Shutty.

The first idea was actually a tool to search your own audio and video recordings. That pivoted into something more useful.

They realized the real money was in the underlying engine that turns speech into text. Deepgram went through Y Combinator in early 2016.

The bet was simple. Old speech recognition was built on clunky decades-old methods.

Deep learning could do it better. They were early and they were right.

Companies pay Deepgram to turn audio into text and text into audio. That is basically it.

A call center wants to transcribe thousands of customer calls. A startup wants its app to understand voice commands.

A company building a phone-answering AI agent needs speech recognition that responds in a fraction of a second. Deepgram sells that as an API.

Developers plug it in and pay based on how much audio they run through it. The pitch to enterprises is speed, accuracy, and price.

It runs cheaper and faster than the big cloud providers for high-volume work.

The core product is the Nova line of speech-to-text models, built for accuracy and low latency on real-world audio. Aura is the text-to-speech side, giving apps natural-sounding voices.

Together they power voice agents that can listen and talk back in real time. Deepgram also sells tools for understanding audio, like detecting topics and summarizing calls.

The whole thing is delivered as a developer API, with options to run it in a company's own environment for privacy-sensitive customers.

Deepgram went after developers and high-volume enterprise customers instead of consumers. No flashy app.

Just an API that was faster and cheaper than Google or Amazon for people running millions of hours of audio. The company trained its own models from scratch rather than bolting onto someone else's.

That let it tune for real-world messy audio like phone calls and noisy rooms. When the voice AI wave hit in 2023 and 2024, Deepgram was already the infrastructure a lot of new voice agent startups quietly ran on.

Being the pick-and-shovel supplier during a gold rush is a good place to be.

Voice AI turned into a knife fight. ElevenLabs, OpenAI, Google, Amazon, and a wave of funded startups all want the same customers.

Big tech can bundle speech recognition into cloud contracts and basically give it away. Deepgram has to stay meaningfully faster, cheaper, or more accurate to justify existing.

The other problem is that speech models keep getting commoditized. What was a moat in 2020 is a checkbox feature in 2025.

Deepgram has pushed into full voice agents and text-to-speech to stay ahead, but the pressure never lets up.

Deepgram raised roughly $86 million through 2023, with Madrona Venture Group and Tiger Global among the backers. Nvidia's venture arm and Y Combinator were early supporters.

In January 2026 it raised a $130 million Series C that pushed its valuation to $1.3 billion, minting a fresh voice AI unicorn. The backing reflects a bet that voice becomes a default way people talk to software, and that someone has to run the plumbing underneath it.