Research Staff, Voice AI Foundations

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.

The Opportunity

Voice is the most natural modality for human interaction with machines. However, current sequence modeling paradigms based on jointly scaling model and data cannot deliver voice AI capable of universal human interaction. The challenges are rooted in fundamental data problems posed by audio: real-world audio data is scarce and enormously diverse, spanning a vast space of voices, speaking styles, and acoustic conditions. Even if billions of hours of audio were accessible, its inherent high dimensionality creates computational and storage costs that make training and deployment prohibitively expensive at world scale. We believe that entirely new paradigms for audio AI are needed to overcome these challenges and make voice interaction accessible to everyone.

The Role

As a Member of the Research Staff, you will pioneer the development of Latent Space Models (LSMs), a new approach that aims to solve the fundamental data, scale, and cost challenges associated with building robust, contextualized voice AI. Your research will focus on solving one or more of the following problems:

• Build next-generation neural audio codecs that achieve extreme, low bit-rate compression and high fidelity reconstruction across a world-scale corpus of general audio.
• Pioneer steerable generative models that can synthesize the full diversity of human speech from the codec latent representation, from casual conversation to highly emotional expression to complex multi-speaker scenarios with environmental noise and overlapping speech.
• Develop embedding systems that cleanly factorize the codec latent space into interpretable dimensions of speaker, content, style, environment, and channel effects -- enabling precise control over each aspect and the ability to massively amplify an existing seed dataset through “latent recombination”.
• Leverage latent recombination to generate synthetic audio data at previously impossible scales, unlocking joint model and data scaling paradigms for audio. Endeavor to train multimodal speech-to-speech systems that can 1) understand any human irrespective of their demographics, state, or environment and 2) produce empathic, human-like responses that achieve conversational or task-oriented objectives.
• Design model architectures, training schemes, and inference algorithms that are adapted for hardware at the bare metal enabling cost efficient training on billion-hour datasets and powering real-time inference for hundreds of millions of concurrent conversations.

The Challenge

We are seeking researchers who

• See "unsolved" problems as opportunities to pioneer entirely new approaches
• Can identify the one critical experiment that will validate or kill an idea in days, not months
• Have the vision to scale successful proofs-of-concept 100x
• Are obsessed with using AI to automate and amplify your own impact

If you find yourself energized rather than daunted by these expectations—if you're already thinking about five ideas to try while reading this—you might be the researcher we need. This role demands obsession with the problems, creativity in approach, and relentless drive toward elegant, scalable solutions. The technical challenges are immense, but the potential impact is transformative.

It's Important to Us That You Have

• Strong mathematical foundation in statistical learning theory, particularly in areas relevant to self-supervised and multimodal learning
• Deep expertise in foundation model architectures, with an understanding of how to scale training across multiple modalities
• Proven ability to bridge theory and practice—someone who can both derive novel mathematical formulations and implement them efficiently
• Demonstrated ability to build data pipelines that can process and curate massive datasets while maintaining quality and diversity
• Track record of designing controlled experiments that isolate the impact of architectural innovations and validate theoretical insights
• Experience optimizing models for real-world deployment, including knowledge of hardware constraints and efficiency techniques
• History of open-source contributions or research publications that have advanced the state of the art in speech/language AI

Ideal Researcher Profile

An ideal researcher, as evidenced by the recurring themes across these foundational papers, excels in five key areas: (1) Statistical & Mathematical Foundations, (2) Algorithmic Innovation & Implementation, (3) Data-Driven & Scalable Systems, (4) Hardware & Systems Understanding, and (5) Rigorous Experimental Design. Below is a synthesis of how each paper highlights these qualities, with references illustrating why they matter for building robust, impactful deep learning models.

1. Statistical & Mathematical Foundations

Mastery of Core Concepts

Many papers, like Scaling Laws for Neural Language Models and Neural Discrete Representation Learning (VQ-VAE), reflect the importance of power-law analyses, derivation of novel losses, or adaptation of fundamental equations (e.g., in VQ-VAE's commitment loss or rectified flows in Scaling Rectified Flow Transformers). Such mathematical grounding clarifies why models converge or suffer collapse.

Combining Existing Theories in Novel Ways

Papers such as Moshi (combining text modeling, audio codecs, and hierarchical generative modeling) and Finite Scalar Quantization (FSQ's adaptation of classic scalar quantization to replace vector-quantized representations) show how reusing but reimagining known techniques can yield breakthroughs. Many references (e.g., the structured state-space duality in Transformers are SSMs) underscore how unifying previously separate research lines can reveal powerful algorithmic or theoretical insights.

Logical Reasoning and Assumption Testing

Across all papers—particularly in the problem statements of Whisper or Rectified Flow Transformers—the authors present assumptions (e.g., "scaling data leads to zero-shot robustness" or "straight-line noise injection improves sample efficiency") and systematically verify them with thorough empirical results. An ideal researcher similarly grounds new ideas in well-formed, testable hypotheses.

2. Algorithmic Innovation & Implementation

Creative Solutions to Known Bottlenecks

Each paper puts forth a unique algorithmic contribution—Rectified Flow Transformers redefines standard diffusion paths, FSQ proposes simpler scalar quantizations contrasted with VQ, phi-3 mini relies on curated data and blocksparse attention, and Mamba-2 merges SSM speed with attention concepts.

Turning Theory into Practice

Whether it's the direct preference optimization (DPO) for alignment in phi-3 or the residual vector quantization in SoundStream, these works show that bridging design insights with implementable prototypes is essential.

Clear Impact Through Prototypes & Open-Source

Many references (Whisper, neural discrete representation learning, Mamba-2) highlight releasing code or pretrained models, enabling the broader community to replicate and build upon new methods. This premise of collaboration fosters faster progress.

3. Data-Driven & Scalable Systems

Emphasis on Large-Scale Data and Efficient Pipelines

Papers such as Robust Speech Recognition via Large-Scale Weak Supervision (Whisper) and BASE TTS demonstrate that collecting and processing hundreds of thousands of hours of real-world audio can unlock new capabilities in zero-shot or low-resource domains. Meanwhile, phi-3 Technical Report shows that filtering and curating data at scale (e.g., "data optimal regime") can yield high performance even in smaller models.

Strategic Use of Data for Staged Training

A recurring strategy is to vary sources of data or the order of tasks. Whisper trains on multilingual tasks, BASE TTS uses subsets/stages for pretraining on speech tokens, and phi-3 deploys multiple training phases (web data, then synthetic data). This systematic approach to data underscores how an ideal researcher designs training curricula and data filtering protocols for maximum performance.

4. Hardware & Systems Understanding

Efficient Implementations at Scale

Many works illustrate how researchers tune architectures for modern accelerators: the In-Datacenter TPU paper exemplifies domain-specific hardware design for dense matrix multiplications, while phi-3 leverages blocksparse attention and custom Triton kernels to run advanced LLMs on resource-limited devices.

Real-Time & On-Device Constraints

SoundStream shows how to compress audio in real time on a smartphone CPU, demonstrating that knowledge of hardware constraints (latency, limited memory) drives design choices. Similarly, Moshi's low-latency streaming TTS and phi-3-mini's phone-based inference highlight that an ideal researcher must adapt algorithms to resource limits while maintaining robustness.

Architectural & Optimization Details

Papers like Mamba-2 in Transformers are SSMs and the In-Datacenter TPU work show how exploiting specialized matrix decomposition, custom memory hierarchies, or quantization approaches can lead to breakthroughs in speed or energy efficiency.

5. Rigorous Experimental Design

Controlled Comparisons & Ablations

Nearly all papers—Whisper, FSQ, Mamba-2, BASE TTS—use systematic ablations to isolate the impact of individual components (e.g., ablation on vector-quantization vs. scalar quantization in