Working at Deepgram in 2026: Voice AI, Culture & What Employees Say

Most people in AI are chasing text. Deepgram bet everything on sound. Founded in 2015 by Scott Stephenson, a former particle physicist who worked on the ATLAS experiment at CERN, Deepgram set out to build speech recognition from the ground up — not by tweaking existing models, but by training custom deep neural networks on raw audio data. A decade later, that bet has paid off: a $1.3 billion valuation, partnerships with IBM and Twilio, and over 200,000 developers building on their APIs.

But what’s it actually like to work there? With a 4.6 Glassdoor rating — one of the highest among the 118 companies in our Culture Directory — Deepgram stands out on paper. We dug into employee reviews, the engineering blog, and the company’s recent trajectory to give you an honest picture of Deepgram as an employer in 2026.

Deepgram at a Glance

Deepgram occupies a fascinating niche in the AI landscape. While OpenAI and Anthropic race to build general-purpose reasoning, and ElevenLabs focuses on voice generation, Deepgram has planted its flag in the infrastructure layer: real-time speech understanding at enterprise scale. Their Nova-3 models achieve word error rates of 5.26% on general English, and their Voice Agent API is the only enterprise-grade platform that handles speech-to-text, text-to-speech, and conversational AI in a single streaming connection. More than 1,300 organizations use their APIs in production.

The Culture: Physicist Founders, Engineer-Led Decisions

Deepgram’s culture is shaped by its origin story. Scott Stephenson didn’t come from a startup accelerator — he came from experimental physics, where you design instruments to observe phenomena that no one else can see. That sensibility permeates the company. Deepgram doesn’t use off-the-shelf models and fine-tune them. They build speech models from scratch, train on proprietary datasets, and optimize inference pipelines down to the hardware level.

This translates into an engineering-driven culture where technical excellence is the primary currency. Employees consistently describe working alongside “talented, humble, and kind people who trust you to do great work.” There’s a conspicuous absence of bureaucratic overhead — at ~260 people, the company is small enough that individual engineers can trace a direct line between their code and the customers it serves.

The learning opportunities are a recurring theme in reviews. Deepgram is doing genuinely novel work at the intersection of deep learning, audio signal processing, and real-time systems. Engineers regularly encounter problems that don’t have textbook solutions — how to reduce inference latency below 300ms while maintaining accuracy across 36+ languages, how to build speech-to-speech models that preserve emotional tone without converting to text at any intermediate step. If you’re the kind of engineer who thrives on hard, open-ended problems, that’s a powerful draw.

But the culture isn’t without tension. Multiple reviewers note that Deepgram operates with a CEO-driven decision-making style. Scott Stephenson is deeply involved in technical and strategic choices, and several employees report that “top-down decisions can override team consensus.” For engineers who thrive with strong founder leadership — think Cursor or early Stripe — this can feel energizing. For those who expect collaborative decision-making at every level, it can be frustrating.

Glassdoor Ratings Breakdown

Deepgram’s 4.6 overall rating places it in the top 10% of our directory, alongside companies like Cursor (4.7) and Linear (4.8). With 46 reviews, it’s a smaller sample than companies like Stripe (2,800 reviews) or Datadog (1,200+), but the consistency across reviews is striking.

The standout here is Career Opportunities at 4.5 — unusually high for a company of this size. This makes sense when you consider the learning velocity. At a 260-person company building novel AI infrastructure, engineers take on scope that would require three levels of seniority at a larger org. The Compensation & Benefits score at 4.5 reflects competitive startup comp with meaningful equity upside at a unicorn valuation. The Work-Life Balance of 4.3 is strong, though some reviews caveat that “you’ll work very hard — it’s hard to find a good balance at times.”

What Employees Actually Say

What employees love

The thread connecting every positive review is trust. Deepgram gives engineers real ownership with minimal gatekeeping. That’s partly structural — at 260 people, there simply aren’t enough layers of management to create bottlenecks — and partly cultural. The founding team came from research labs where intellectual autonomy is the default, and that norm has persisted as the company scaled.

What could be better

The cons cluster around two themes: (1) the intensity that comes with a fast-growing startup in a competitive market, and (2) communication gaps that some attribute to the CEO-driven decision style. The communication issue is worth noting. Several reviewers describe situations where strategic context wasn’t shared broadly enough, leaving engineers to build without understanding the full picture. At a company this size, that’s fixable — and likely something the leadership is actively working on as they scale from 150 to 260+ people.

Compensation & Benefits

Deepgram’s compensation is competitive for a Series C startup, particularly given the equity upside at a $1.3 billion valuation. The 4.5 Glassdoor rating for Compensation & Benefits puts it in the same tier as Stripe (4.5) and Anthropic.

Based on verified employee-reported salary data, total compensation for software engineers at Deepgram ranges from approximately $210K to $300K+, depending on level and specialization. The median engineering package sits around $266K, which includes base salary, equity, and bonuses. That’s below frontier AI labs like OpenAI ($350K–$550K+) and Anthropic ($300K–$490K), but competitive for a sub-300-person startup — and the equity component carries real upside given the $1.3B valuation and growing enterprise revenue.

The $130M Series C round, led by AVP with participation from BlackRock, Twilio, ServiceNow Ventures, SAP, and Citi Ventures, adds institutional credibility. These aren’t speculative bets from early-stage VCs — they’re strategic investments from companies that are themselves customers of Deepgram’s voice AI infrastructure. That investor profile matters for equity value: enterprise customers as investors typically signal durable revenue, not hype-driven valuation.

Engineering Culture & Tech Stack

Deepgram’s engineering is what differentiates it from nearly every other company in the voice AI space. While competitors often wrap OpenAI’s Whisper model in an API layer, Deepgram trains its own speech models from scratch. That’s a harder path — it requires building custom training infrastructure, assembling proprietary datasets, and optimizing every layer of the inference pipeline — but it produces models that are faster, more accurate, and more customizable for enterprise use cases.

Tech Stack

The core inference engine is written in Rust and C++ for maximum performance — when you’re processing real-time audio at sub-300ms latency, every microsecond in the pipeline matters. Python drives the ML training pipelines and research experimentation, while Go powers the backend API services. CUDA and custom GPU kernels optimize the deep learning inference path. This is not a “wrap a model in FastAPI” stack — it’s systems-level engineering that spans from GPU memory management to WebSocket streaming protocols.

What engineers work on

• Nova model family. Deepgram’s flagship speech-to-text models, now in their third generation. Nova-3 achieves state-of-the-art accuracy across 36+ languages. The ML team works on architecture design, training data curation, and inference optimization.
• Speech-to-speech models. A milestone achievement: end-to-end models that translate speech directly without intermediate text conversion, preserving nuance, intonation, and emotional tone. This is frontier research, not incremental improvement.
• Voice Agent API. The industry’s first enterprise-grade, real-time conversational AI API. Engineers build the streaming infrastructure that handles STT, TTS, and orchestration in a single connection — replacing what previously required stitching three separate services together.
• On-premise deployments. For customers in healthcare, finance, and government who can’t send audio to the cloud. Engineers work on containerized deployments that run Deepgram’s full model stack in customer environments.

The technical depth is genuine. Deepgram has processed over 50,000 years of audio and transcribed more than one trillion words. Their IBM partnership — integrating Deepgram’s speech capabilities into watsonx Orchestrate as IBM’s first voice partner — signals the kind of enterprise trust that only comes from battle-tested infrastructure.

Recent Momentum: The $130M Series C

Deepgram entered 2026 with significant momentum. The January 2026 Series C round raised $130M at a $1.3B valuation, making Deepgram a unicorn. But the funding story is less about the number and more about what it signals.

First, Deepgram reached the round cash-flow positive — meaning the $130M isn’t survival capital but growth fuel. Annual usage had grown 3.3x over the preceding four years, driven by the explosion of voice AI agents, call center automation, and real-time transcription in sectors like healthcare and legal. Second, the investor mix (BlackRock, Stanford, Columbia, plus strategic partners Twilio, SAP, ServiceNow, and Citi Ventures) reflects institutional confidence in the durability of the business, not just the technology.

The funding is earmarked for three priorities: expanding the team from 150 to 250+ employees (mostly engineering and sales), launching the “Powered by Deepgram” program for OEM partnerships, and building a Voice AI Collaboration Hub in San Francisco. They also acquired OfOne, a Y Combinator-backed startup specializing in voice AI for restaurants — an early move into vertical-specific voice applications.

Who Thrives at Deepgram

Based on the culture signals, compensation data, and employee feedback, here’s who tends to do well:

• Systems-level engineers. If you think in terms of memory layouts, latency budgets, and GPU utilization, this is your kind of company. Deepgram’s engineering problems are closer to what you’d find at Cerebras or a robotics company than at a typical SaaS startup.
• ML researchers who want production impact. Unlike pure research labs, every model improvement at Deepgram ships to production and serves real customers within weeks. If you want your research to matter beyond a paper, that’s a compelling proposition.
• Self-starters who don’t need hand-holding. The “few guardrails” approach means you get autonomy, but it also means you need to be comfortable navigating ambiguity. Deepgram won’t give you a detailed spec — they’ll give you a problem and trust you to figure it out.
• People excited about the voice AI market. This matters more than it sounds. Voice AI is having its ChatGPT moment — customer service bots, medical dictation, real-time translation, AI phone agents. If you believe voice is the next major interface layer (and the $130M from strategic investors suggests the market does), Deepgram is the infrastructure play.

Deepgram is not ideal for engineers who want highly structured environments with clear career ladders and extensive documentation. It’s a fast-moving startup that prioritizes shipping over process. If you want the career ladder transparency of HubSpot or the writing-culture structure of Stripe, this will feel chaotic. If you want the flat, high-trust, high-impact energy of a company building something genuinely new, it’s one of the best places in AI to do it.

Open Positions at Deepgram

Deepgram currently has 65 open positions listed on our platform, spanning ML engineering, backend infrastructure, developer relations, and enterprise sales. With the Series C funding specifically designated for team expansion, 2026 is an active hiring window. Roles span San Francisco, Ann Arbor (where the company has deep ties to the University of Michigan research community), and remote.

For full details on Deepgram’s open roles, culture values, and side-by-side comparisons with other companies, visit the Deepgram culture profile page.