A research team developed NeuroBuds, an earphone-based sensor, and an AI model that learns reusable patterns from everyday physiological signals around the ear, supporting gaze estimation and tasks spanning all five senses.

Most earphones are designed to deliver sound. NeuroBuds does the opposite: it listens to the body. Researchers from KAIST and collaborating institutions developed an earphone-based sensing prototype and an AI method that learns from tiny electrical signals around the ear. The work turns a familiar wearable into a potential platform for understanding gaze and responses related to the five senses in everyday settings.

Electrical activity from the brain, eyes, and facial muscles contains useful information about human behavior and physiology. Eye movements, for example, create slow signal changes that can help estimate where a person is looking, while faster patterns can reflect muscle or neural activity. Conventional measurements, however, usually require bulky headsets and controlled laboratories. Existing AI systems are also typically built for one task at a time, with different signal filters and model designs for gaze, emotion, sleep, or movement.

NeuroBuds addresses the data problem with a lightweight earhook-style device that records signals near both ears (Figure 1). Using the prototype, the team created DailySense. The dataset contains 50 hours of unlabeled recordings from 22 participants going about ordinary activities such as walking, eating, and talking. It also includes 20 hours of labeled data across six tasks: gaze estimation, interest while watching video, interest while listening to audio, and classification of taste, surface texture, and smell. Together, these tasks span sight, hearing, taste, touch, and smell.

To avoid building a separate model for every task, the researchers introduced Physiology-informed Multi-band Tokenization. Rather than treating the recording as one mixed waveform, the method separates it into 12 biologically meaningful frequency ranges, much like isolating different sections of an orchestra.

Short segments from each range become tokens that the model can learn from. By first reconstructing everyday recordings without manual labels, the model learns reusable patterns and later identifies which frequency ranges matter for a new task.

On DailySense, this pretraining raised the average classification score from 81.9% to 87.6% and reduced gaze error to 6.0 degrees, as shown in Figure 2. The method also outperformed strong baselines on four public datasets covering emotion recognition, sleep stages, and imagined movement. Tests on a commercial smartphone further showed that real-time processing is feasible, with an average inference time of 25 milliseconds.

The study shows how earphones could evolve from audio accessories into general-purpose physiological sensing platforms. Potential applications include hands-free interfaces, context-aware wearables, and personalized monitoring, although broader and more diverse studies are still needed before deployment. The work was presented at the International Conference on Learning Representations (ICLR) 2026.