Skip to content
HOMETECH DIRECTORYNEWS & DEVELOPMENTSRELOCATEEMPLOYERSNEIGHBORHOODSABOUTSchedule a Call

Deepgram

Deepgram is a voice AI and speech recognition company founded by U-M physicists. HQ San Francisco, with an Ann Arbor engineering satellite and deep U-M...

A2Hub Editorial· Rusty Roof Media· Updated May 4, 2026

Deepgram

Deepgram is one of the most consequential voice AI companies to come out of the University of Michigan research ecosystem, even though its corporate headquarters sits on Market Street in San Francisco. The company was founded in 2015 by two U-M physics alumni whose earliest prototype grew out of signal-processing work in a Michigan lab, and the Ann Arbor thread has shaped its engineering culture and hiring pipeline from the start. Deepgram builds end-to-end deep learning speech recognition and voice AI models used by tens of thousands of developers and enterprise teams. It has raised roughly $86M in disclosed venture funding across Y Combinator, Tiger Global, NVIDIA, Wing Venture Capital, Citi Ventures, and Madrona. For Ann Arbor, Deepgram sits in the diaspora column. A company originated in town, scaled on the coast, and still drawing meaningfully on Michigan-trained engineers.

Quick facts

  • Founded: 2015
  • Founders: Scott Stephenson (CEO), Noah Shutty
  • HQ: 548 Market Street, San Francisco, California
  • Ann Arbor presence: Satellite engineering footprint; small, tight-knit team historically anchored by founder ties to U-M
  • Headcount: Approximately 180 to 220 globally as of latest disclosures; specific Ann Arbor team size not publicly broken out
  • Funding: Approximately $86M total disclosed across seed through Series C
  • Key investors: Tiger Global Management, Wing Venture Capital, NVIDIA, Citi Ventures, Y Combinator, Madrona Venture Group, NEA
  • Current leader: Scott Stephenson, CEO and co-founder
  • Sector tags: AI, speech recognition, voice AI, developer tools
  • Careers: deepgram.com/careers
  • Website: deepgram.com
  • LinkedIn: linkedin.com/company/deepgram

What Deepgram makes

Deepgram builds foundation models for speech and voice, packaged as developer APIs and deployable enterprise endpoints.

The core product line is automatic speech recognition. Deepgram's approach from the beginning was end-to-end deep learning trained on raw audio rather than the older pipeline of acoustic models, phoneme dictionaries, and language models stitched together. That architectural call, which was unconventional in 2015, is what let the company iterate faster than incumbents as transformer-era speech modeling matured. The current flagship model family, Nova, is marketed as faster and more accurate than comparable offerings from Google, AWS, and Azure on developer-facing benchmarks.

Around the transcription core, Deepgram has built an adjacent voice AI stack. That includes speech-to-text for streaming and batch audio, text-to-speech with the Aura model family, a voice agent API for conversational applications, and specialized models for medical, contact-center, and broadcast use cases. Customers integrate Deepgram directly via REST and WebSocket APIs or deploy the models on-premise for regulated environments.

The business model is usage-based API pricing with enterprise contracts on top. Customers include contact-center platforms, health-tech companies transcribing clinical visits, media monitoring firms, and several consumer voice applications that would not be comfortable naming a vendor publicly. NVIDIA's position as both investor and infrastructure partner is worth noting. Deepgram's inference stack is heavily optimized for GPU deployment, and the NVIDIA partnership extends to reference architectures and co-marketing around enterprise voice AI.

Developer posture is a first-class commitment. The company publishes model cards, keeps documentation current, runs an active developer advocate function, and has historically released open-source tooling around the API. That posture is what built bottom-up adoption inside engineering teams before the enterprise sales motion matured.

The Ann Arbor story

Deepgram's origin story runs through the Michigan physics department. Scott Stephenson and Noah Shutty were both U-M physics graduates. Stephenson was working on a particle physics PhD that involved processing massive volumes of signal data. Parsing rare-event signal out of noisy time-series is, as Stephenson has observed in interviews, not so different from the problem of pulling structured meaning out of raw audio. The early Deepgram prototype grew directly out of that sensibility.

The founders went through Y Combinator in the winter 2016 cohort. From there the company relocated its corporate center of gravity to the Bay Area, where the investor pool and enterprise customer base for AI infrastructure sat. But the Ann Arbor thread never fully disappeared. Deepgram retained a Michigan engineering satellite and described it on the company's own careers material as a tight-knit team. Hires have consistently included U-M computer science and electrical engineering graduates, and the founder network has stayed active in Ann Arbor AI conversations through the Voxel51 and Clinc community over the past decade.

For the Ann Arbor ecosystem, Deepgram is best framed as diaspora rather than local anchor. The company is not an Ann Arbor employer at the scale of Duo Security or Censys, and recruiters should not approach it expecting a full in-town engineering organization. What it offers instead is a credible remote-plus-satellite option for Michigan-trained engineers who want to work on voice AI at the research frontier without relocating to California.

The broader pattern is worth naming. A significant share of University of Michigan-originated AI companies make a similar move in their growth years. The corporate headquarters ends up coastal, the investor syndicate is mostly coastal, but the engineering talent draw keeps looping back through Ann Arbor. Deepgram is one of the clearest examples of that pattern, which is why it earns a profile on this site alongside the Ann Arbor AI companies that never left.

Funding and exits

Deepgram's disclosed funding history as reported across PitchBook, Crunchbase, and company announcements:

  • Seed and Y Combinator (2016): Winter 2016 cohort, followed by a seed round of approximately $1.8M
  • Series A (2018): Approximately $12M led by Wing Venture Capital
  • Series B (December 2020): Approximately $25M led by Tiger Global with NVIDIA, Wing, Y Combinator, and other participants
  • Series B extension (2021): Additional capital at higher valuation, roughly $47M total for the B round per reported aggregates
  • Series C (2022): Approximately $72M led by Madrona Venture Group with participation from Citi Ventures, NVIDIA, Wing, and Tiger Global

Cumulative reported funding sits near $86M. Some press reports cite a higher total when including convertible instruments. The company has not publicly raised a priced round since the 2022 Series C as of early 2026. Any extension or Series D activity should be verified against recent filings before publication.

There has been no exit. Deepgram remains private. Public commentary from Stephenson has framed the company as building toward a large standalone voice AI business rather than optimizing for an acquisition. Given the current strategic posture of voice AI inside the hyperscalers and NVIDIA's investor position, an IPO path is more plausible than an acquisition, though neither is imminent on public timelines.

Culture and hiring

Deepgram hires machine learning research engineers, infrastructure engineers, developer advocates, and enterprise solutions engineers. The work leans heavily toward large-model training, inference optimization, and production voice systems at scale. The team profile is technically deep, with a high concentration of graduate-trained machine learning and systems engineers.

Work model is remote-first with hub concentrations in San Francisco and the Ann Arbor satellite. Compensation is competitive with Bay Area AI-infrastructure companies at the same stage. Equity is private-company stock with standard four-year vesting. Interview loops are rigorous and technically specific, with a pronounced emphasis on deep-learning systems fundamentals and applied signal processing.

Public employee signal on Glassdoor and Blind is limited in volume but consistently points to strong engineering peers, demanding work, and a culture that is less cushioned than larger hyperscaler research organizations. That is consistent with the company's stage and with founder posture.

Why they matter for Ann Arbor tech

Deepgram matters for Ann Arbor for two reasons. First, it is proof that U-M physics and computer science training can originate foundation-model-scale AI companies. The founder story, signal processing to speech recognition, is precisely the kind of interdisciplinary arc the region likes to claim. Second, the satellite engineering presence and the hiring pipeline keep Ann Arbor in the mix for ambitious ML engineers who do not want to relocate. For the broader Ann Arbor AI pillar, Deepgram stands next to Voxel51 and Clinc as a U-M-originated company that reached scale. Unlike those two, it chose coastal headquarters, which is an honest part of the record.

FAQ

Where is Deepgram actually headquartered? Deepgram's corporate headquarters is at 548 Market Street in San Francisco, California. The company operates a small Ann Arbor engineering satellite tied to its founder network at the University of Michigan.

Who founded Deepgram, and what is the Ann Arbor connection? Scott Stephenson (CEO) and Noah Shutty co-founded Deepgram in 2015. Both are University of Michigan physics graduates. Stephenson was working on a particle physics PhD that involved processing large audio and sensor data volumes during his U-M years. That signal-processing background translated directly into the company's end-to-end deep learning approach to speech recognition.

What does Deepgram build? Deepgram builds foundation models for speech and voice, including automatic speech recognition (the Nova model family), text-to-speech (Aura), a voice agent API for conversational applications, and specialized models for contact-center, medical, and broadcast use cases. Products are delivered via cloud APIs and on-premise deployment.

How much has Deepgram raised? Approximately $86M in disclosed venture funding across seed, Series A, B, and C rounds, with lead investors Wing Venture Capital, Tiger Global, and Madrona Venture Group. NVIDIA participated as both strategic investor and infrastructure partner.

Does Deepgram hire in Ann Arbor? Yes, though at a smaller scale than its San Francisco hub. The Ann Arbor engineering satellite is described on Deepgram's own materials as a tight-knit team. The company also hires remotely across the United States, which keeps Michigan-based engineers in the candidate pool for most roles.

Who competes with Deepgram? On speech recognition, Deepgram competes with AWS Transcribe, Google Cloud Speech-to-Text, Azure Speech, OpenAI Whisper (open source), AssemblyAI, and Rev.ai. Its positioning emphasizes latency, accuracy on challenging audio, and developer experience over the cloud hyperscaler incumbents.

Related Ann Arbor AI companies

  • Voxel51 is the Ann Arbor headquartered computer vision infrastructure company with a similar U-M origin pattern.
  • Clinc is a conversational AI spinout from U-M's EECS department, with products aimed at banking.
  • MemryX builds edge AI accelerator chips out of Ann Arbor and shares the U-M deep-tech lineage.
  • The Ann Arbor AI companies hub covers the full regional roster.
  • The Ann Arbor tech directory is the site homepage.