At Blynt, we are building the interface of the future: voice that works. Our team previously built Sonos Voice Control, delivering a production-grade voice experience to millions of users worldwide while pushing the technical boundaries, with a fraction of the resources of Big Tech and running the entire AI stack directly on Sonos speakers under extreme runtime and optimization constraints.

We focus on everything that happens between the microphone and the AI agent: turning raw, continuous audio into structured, real-time events that downstream systems can act on. Our model stack spans real-time transcription and turn-taking, with environmental and paralinguistic understanding on the roadmap. Powering the next generation of real-world agentic interactions, from robots to wearables and beyond, requires a step change in the underlying technology. That's what we're building.

Our team is remote first, with two hubs in Annecy and Paris. You can work remotely from anywhere in 🇫🇷 or 🇨🇭 if that suits you. If you get your energy from being around other people, the hubs are open and used. Both are normal here, and nobody has to justify their choice. We also get everyone together in person once a month either in Annecy or Paris.

The role

As a Research Engineer in Audio & Speech, you will own key parts of our model stack from research through production. This is not a pure research role where papers are the output. You will design, train, evaluate and deploy models that run in real-time, in production, at scale. You'll work at the intersection of research and engineering, translating state-of-the-art ideas into systems that work reliably in the real world.

What you will work on

• Model development: Design, train, and iterate on speech and audio models across Audio LLMs, Speech LLMs, Full-duplex Turn taking, Speaker Identification and Diarization
• Domain adaptation & personalization: Develop scalable techniques for fine-grained customization of speech models to specific acoustic environments, demographics, and vocabularies, enabling per-customer accuracy that off-the-shelf models can't deliver.
• Evaluation & benchmarking: Build rigorous evaluation pipelines that go beyond standard benchmarks. Design metrics and test sets that reflect the real-world conditions our customers operate in.
• Production deployment: Your models don't live in notebooks. They run in real-time streaming pipelines with hard latency constraints. You care about inference efficiency, model size, and how your architecture choices affect the system end-to-end.

What we're looking for

First and foremost, nice human beings who are genuinely excited by what we are trying to do, which is to make a real dent in the Conversational AI landscape. The list below is not a checklist. Very few candidates will tick every line, and we would rather hear from a wide range of backgrounds and experiences.

• PhD in Machine Learning or related field, or MSc with several years of hands-on experience building production-grade Audio/Speech systems
• Proficient with PyTorch (or equivalent deep learning framework)
• Deep experience in at least one of the following: Neural Speech Enhancement, Keyword Spotting, Automatic Speech Recognition, Audio Classification, Audio LLMs or Speech Synthesis
• Comfortable writing production Python code, not just research prototypes

Nice to have: