Italian researcher at DTU tests small AI specialist models for Bang & Olufsen headphones

Postdoc Alessandro Cerioli at DTU is testing whether headphones work better with a team of small AI specialist models than with one large general model. He is part of REINS, a DIREC project working with Bang & Olufsen. Cerioli is Italian, trained as a computer engineer at Politecnico di Milano, and came to Denmark for a master's in sound and music computing at Aalborg University. He later worked as an embedded developer at the hearing company GN and did an industrial PhD with Jabra and DTU.

The problem is that headphones must enhance speech as the wearer moves from a quiet office to a busy street, react instantly, and last all day on a tiny battery. A large general model can handle any situation but is too big to fit in an earbud, and too slow if every sound has to be sent to a data center and back. Cerioli calls the large option the "global solution": less expert, but covering everything. The alternative is several compact TinyML models, each trained as an expert for one context. As he puts it, the more you improve quality, the more hardware resources you usually need, so the task is finding the best trade-off.

He says the question goes beyond headphones. Most of today's AI is built big and uses a great deal of energy, and he considers it unsustainable financially and in terms of energy and water. He argues that AI could be used in many more places, closer to where things happen, with far less energy. Working out where going big pays off and where going small is enough is the first step.

The article defines three terms. Edge AI runs directly on a device such as earbuds or a sensor, which saves energy, removes network delay and works without an internet connection. TinyML is AI on very small, low-power microcontrollers, often with only kilobytes of memory. Adaptive AI changes how it works as conditions change, for example by switching to a different specialised model when the language, noise or task changes.

REINS runs from 2025 to 2027 and aims to make AI run efficiently on devices with very little computing power. Its partners are DTU, the University of Southern Denmark, the Alexandra Institute, Bang & Olufsen and Leica Geosystems.