A Wearable Ultra-Low-Power System for EEG-Based Speech-Imagery Interfaces
Abstract
Speech imagery—mentally simulating speech without vocalization—is a promising control modality for brain–computer interfaces. This paper delivers the first end-to-end demonstration of EEG-based speech-imagery decoding on a low-channel, ultra-low-power wearable device. Building on the BioGAP platform and the lightweight VOWELNET neural network, the system classifies 13 classes (vowels, commands, and rest) with subject-specific training, reaching up to 50% accuracy (42.8% average) while operating in real time at 25.93 mW on a GAP9 processor. The device runs continuously for more than 21 hours with 40.9 ms latency and explores continual learning strategies and electrode placement.
Key Highlights
- First fully wearable EEG speech-imagery decoding system supporting a 13-class vocabulary on the BioGAP platform.
- Lightweight VOWELNET network delivers up to 50% accuracy (42.8% average) with 40.9 ms latency at 25.93 mW on GAP9.
- Operates for more than 21 hours and explores electrode placement and continual learning strategies for practical deployment.
Thorir Mar Ingolfsson
Postdoctoral Researcher
I develop efficient machine learning systems for biomedical wearables that operate under extreme resource constraints. My work bridges foundation models, neural architecture design, and edge deployment to enable real-time biosignal analysis on microwatt-scale devices.