A Wearable Ultra-Low-Power System for EEG-based Speech-Imagery Interfaces.

Abstract

Speech imagery-the process of mentally simulating speech without vocalization-is a promising approach for brain-computer interfaces (BCIs), enabling assistive communication for individuals with speech impairments or to enhance privacy. However, existing EEG-based speech imagery systems remain impractical for use outside specialized laboratories due to their reliance on high-channel-count and resource-intensive machine learning models running on external computing platforms. In this work, we present the first end-to-end demonstration of EEG-based speech imagery decoding on a low-channel, ultra-low-power wearable device. Building on our previous work on vowel imagery, we introduce an extended framework leveraging the BioGAP platform and VOWELNET, a lightweight neural network optimized for embedded speech imagery classification. In particular, we demonstrate state-of- the-art accuracy in the classification of an expanded vocabulary comprising vowels, commands, and rest states (13 classes) with a subject-specific training approach, achieving up to 50.0% for one subject (42.8% average) in multi-class classification. We deploy our model on an embedded biosignal acquisition and processing platform (BioGAP), based on the GAP9 processor, for real-time inference with minimal power consumption (25.93 mW). Our system achieves continuous execution for more than 21 hours on a small LiPo battery while maintaining classification latencies of 40.9 ms. Finally, we also explore the benefits of applying Continual Learning techniques to progressively improve the system's performance throughout its operational lifetime, and we demonstrate that electrodes located on the temporal area contribute the most to the overall accuracy. This work marks a significant step toward practical, real-time, and unobtrusive speech imagery BCIs, unlocking new opportunities for covert communication and assistive technologies.