Research Staff, Voice AI Foundations [UK - Remote]

Deepgram operates a platform that powers the emerging voice AI economy, providing real-time APIs for speech-to-text, text-to-speech, and production-grade voice agents at scale. The company has processed over 50,000 years of audio and transcribed more than 1 trillion words, working with more than 200,000 developers and 1,300+ organizations including Twilio, Cloudflare, and Sierra. Deepgram's voice-native foundation models deliver unmatched accuracy, low latency, and cost efficiency through cloud APIs and on-premises deployment.

Deepgram operates in an AI-first environment where experimentation and active use of advanced AI tools are core to daily work. The company moves at the pace of AI, meaning your role will evolve rapidly alongside the field. You should expect your work to be dynamic, unpredictable, and deeply integrated with the latest AI capabilities.

This Research Staff position focuses on pioneering Latent Space Models as a fundamental new approach to voice AI. Your research tackles core problems in audio: real-world data scarcity, acoustic diversity, and the computational costs of training and deployment at scale. You will work on one or more critical areas: neural audio codecs that compress audio while preserving fidelity, steerable generative models for diverse human speech synthesis, embedding systems that factorize codec latent space into speaker, content, style, and environment dimensions, and systems that leverage latent recombination to generate synthetic audio at unprecedented scale. Your work will also include designing model architectures and inference algorithms optimized for hardware efficiency, enabling training on billion-hour datasets and real-time inference for hundreds of millions of concurrent conversations.

What you'll do

• Develop next-generation neural audio codecs achieving extreme low bit-rate compression with high-fidelity reconstruction across diverse global audio
• Design steerable generative models that synthesize the full spectrum of human speech, from casual conversation to emotional expression to multi-speaker scenarios with noise
• Create embedding systems that decompose latent space into interpretable, controllable dimensions for speaker, content, style, and environment
• Generate synthetic audio at scale through latent recombination techniques, unlocking new data scaling paradigms
• Develop hardware-optimized architectures and inference algorithms for cost-efficient training and real-time deployment
• Design controlled experiments that validate or invalidate ideas rapidly, scaling successful proofs-of-concept 100x over

What you'll bring

• Strong mathematical foundation in statistical learning theory, with expertise in self-supervised and multimodal learning
• Deep knowledge of foundation model architectures and multi-modal scaling across training
• Proven ability to move between theory and implementation, deriving novel mathematical formulations and building them efficiently
• Experience building data pipelines that process and curate massive datasets while preserving quality and diversity
• Track record designing controlled experiments that isolate architectural impact and validate theoretical insights
• Experience optimizing models for real-world deployment, understanding hardware constraints and efficiency techniques
• Comfort with rapid change, experimentation, and continuous learning in an AI-driven environment

Nice to have

• Background in audio processing, speech synthesis, or speech recognition
• Experience deploying models at scale in production environments
• Familiarity with bare-metal optimization and inference acceleration

What they offer

• Remote position based in London, UK
• Full-time employment
• Opportunity to work on foundational research shaping the voice AI industry