About Phonic
Phonic is the speech-to-speech model company. Our mission is to make speaking with agents delightful, and we believe that requires models that hear and answer in native audio. To make that possible in production, we work on every layer of the stack: data, pre-training, post-training, kernels, and infrastructure. We’ve raised over $30M from top-tier investors, including First Round, Lux Capital, and Felicis. Our models power voice agents on six continents for companies like Ashby, Maven AGI, and a Fortune 50 IT infrastructure company, handling everything from customer support and patient intake to recruiting interviews and dispatch.
About the Team
We’re a small team of researchers and engineers from MIT, Stanford, CMU, MosaicML, Citadel, and Genesis Therapeutics, including former founders, an IMO medalist, and an International Linguistics Olympiad medalist. We work in person in San Francisco five days a week. Everyone here is unusually obsessed with our customers: we’ll happily spend hours perfecting the way our model reasons while it talks, or the way it makes a caller feel heard, going as deep into the stack as it takes.
The Role
As a Research Intern at Phonic, you'll work directly on the core problems that make Phonic's voice AI feel genuinely human. You'll own a research direction end-to-end — from identifying the right problem to designing experiments, developing novel methods, and seeing results through to production impact. This isn't a role focused on incremental improvements; we're looking for someone with the taste to identify what matters, the rigor to pursue it correctly, and the drive to ship it. You'll be fully in-person in our SF office.
What You'll Do
Identify high-leverage research problems across the voice AI stack from audio understanding to audio output, and take full ownership of driving them forward
Design and run rigorous experiments that analyze architectural trade-offs to understand how design choices influence a model’s scalability, latency, and quality
Curate massive training datasets and execute rigorous experiments to determine exactly how data quality shapes model behavior and performance
Work directly with research scientists and engineers to move fast from prototype to production
Build the training pipelines, evaluation frameworks, and tooling that let us experiment and iterate quickly
What You'll Bring
A track record of original work: you've found a real problem, developed an approach, and seen it through
Proficiency in PyTorch (or JAX), and the ability to implement models cleanly from papers
Fluency in the math, probability, optimization, and linear algebra underlying model behavior
You move fluidly between ideas and implementation; you don't just think about problems, you build things
Clear, precise written and verbal communication
Nice To Have
Research experience in speech, audio, or language modeling (ASR, TTS, LLMs, codec models)
Familiarity with generative modeling techniques: diffusion, flow matching, or autoregressive models
Experience with RLHF or preference optimization
Competitive programming or olympiad background
Publications or preprints at venues like NeurIPS, ICML, ICLR, Interspeech, ICASSP, or ACL
Benefits
💸 Top-tier compensation: in order to get the best talent, we provide salary and equity that recognize your skillset
🥗 Meals: free breakfast, lunch, and dinner provided in the office
🤝 We have regular off-sites and team celebrations