# Research Staff, Voice AI Foundations — Deepgram

- Location: San Francisco, CA (remote)
- Posted pay: $149k–$338k/yr
- Posted: Aug 23, 2026
- Status: open (checked Nov 21, 2026)
- Apply (employer's site): https://jobs.ashbyhq.com/Deepgram/7cf277bc-a192-4007-a77a-fafbf8276a5e
- Employer record: https://visajobs.io/company/deepgram.md

## Sponsorship history
- Tier: Employer history only
- With this exact title: 0

## Description
Company Overview
Deepgram is bringing the world’s most natural interface, voice, to the technology we use every day. Our speech models and real-time infrastructure power AI systems that keep up with the way people actually speak. More than 5,000 organizations run conversational voice AI on Deepgram in production, including Twilio, Cloudflare, Sierra, Decagon, Vapi, Cresta, Granola, and Jack in the Box.
The Deepgram Voice Agent Platform brings together speech-to-text, text-to-speech and speech-to-speech, with voice-native foundation models available through cloud APIs or self-hosted and on-premises deployments. Built for accuracy, low latency and cost efficiency, Deepgram makes high-quality voice experiences practical to build and run at scale. More than 650,000 developers have used Deepgram to process more than 200,000 years of audio. Every day, more of the world’s conversations with AI run on our models.
Operating Principles
Deepgram is a research company that productizes as fast as possible. Four Operating Principles shape how we think, decide, and work together:
Compress before you communicate. You think it through first, then lead with the bottom line.
Resist entropy. You question what already exists before adding more process or tools.
Solve the category. You fix recurring problems at the source, not one case at a time.
Work at the frontier. You go after problems no one has solved yet.
We look for people who already work this way, and these principles shape how we grow and evaluate our team.
AI-First by Default
We build AI that millions of people talk to, and we run the company the same way. AI is how we centralize work here, and is core to how we operate.
That means more than using the tools. We expect you to:
Build, not just use. Design your own workflows and agents, and use AI to remove whole categories of work instead of doing the same work faster.
Own the output. AI can speed up the work. It doesn’t make anyone less accountable for it. You understand, verify, and stand behind everything you share, including what an agent produces for you.
Spend the time you get back at the frontier. The capacity AI creates goes toward problems no one has solved yet.
Keep pace. Models and tools change every month, and so will your work. You adopt what’s new quickly and drop what no longer earns its place.
We judge AI by what it helps you deliver and whether you stand behind the result, not by how often you use it. In interviews, expect to walk us through a workflow you built and where you drew the line between AI and your own judgment.
Team Overview
Deepgram’s Frontier Research Staff works at the frontier of audio AI, across both perception and generation, from speech-to-text to text-to-speech. We start from open-ended questions about what models can do and how they should be built, then pursue new approaches through experiments that become the research foundations for our STT and TTS systems. We measure success by whether the work advances what our models can do. Real-time is the constraint that shapes every question we ask.
Latent Space Models is our bet on a new paradigm for audio AI. Rather than building on current approaches, we use a learned codec latent as a shared foundation for generation, understanding, and synthetic data, opening up capabilities and a scale those approaches can’t reach. The goal is for voice to become the natural interface to technology.
The team works closely with engineering, infrastructure, product, data, and marketing to carry breakthroughs from first experiment into production.
Key Goals
Own a hard part of the Latent Space Models agenda end to end, from research through a model that ships under real-time constraints
Build next-generation neural audio codecs that compress general audio to extremely low bit rates while reconstructing it with high fidelity
Pioneer steerable generative models that synthesize the full range of human speech from the codec latent (casual conversation, strong emotion, multiple speakers with noise and overlap) with natural prosody and precise control, in real time
Factorize the codec latent into interpretable dimensions such as speaker, content, style, environment, and channel, so the team can steer generation and recombine latents with intent
Use latent recombination to generate synthetic audio at scale, and to train speech-to-speech systems that understand anyone and respond with empathy and purpose
Build durable data, evaluation, and training systems that curate and learn from massive, diverse audio streams without sacrificing quality, so each experiment gets faster than the last
Co-design architectures, training schemes, and inference with bare-metal hardware, keeping training cost-efficient and inference real-time at world scale
Help set research direction for the team, and make complex results clear and actionable for researchers, engineers, product, data, and marketing
Minimum Skills, Knowledge & Capabilities
Deep, demonstrated expertise in generative modeling, such as diffusion, latent models, or controllable generation, in speech, image, video, or another creative generative domain
Strong mathematical grounding in statistical learning, especially self-supervised and multimodal learning, with the ability to derive new formulations and implement them efficiently
Data-, representation-, and eval-centric: you spot problems in the data, interrogate what a model’s latents actually encode, and build evals that tell the truth
Designs controlled experiments and ablations that isolate what works, and drops ideas quickly when the evidence says so
Has built new foundation-model architectures or scaled existing ones across modalities
Works AI-native: designs Human+AI research workflows, and builds your own tooling, infrastructure, or evals to unblock work instead of waiting on them
Reasons from first principles, carries ideas across domains, and explains deep technical work at whatever altitude the listener needs
Advanced proficiency in Pytho

> History, not a promise: past H-1B filings do not mean an employer will sponsor you or a given role. Data: U.S. DOL OFLC LCA disclosure (FY2024–FY2026 for role counts, certified LCAs) and the USCIS H-1B Employer Data Hub. Not legal advice. Methodology: https://visajobs.io/methodology
