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

Area of Science:

• Neuroscience
• Biomedical Engineering
• Computer Science

Background:

• Speech imagery, mentally simulating speech without vocalization, is key for brain-computer interfaces (BCIs).
• Current EEG-based speech imagery systems are limited by high channel counts and resource-intensive models, hindering practical application.
• Existing systems require external computing platforms, making them unsuitable for portable use.

Purpose of the Study:

• To demonstrate the first end-to-end EEG-based speech imagery decoding on a low-channel, ultra-low-power wearable device.
• To develop and deploy a lightweight neural network (VowelNet) optimized for embedded speech imagery classification on the BioGAP platform.
• To achieve real-time, continuous operation with minimal power consumption for practical BCI applications.

Main Methods:

• Developed an extended framework for speech imagery decoding, including vowels, commands, and rest states (13 classes).
• Utilized the BioGAP platform with a GAP9 processor for embedded real-time inference.
• Implemented a subject-specific training approach and explored Continual Learning techniques.

Main Results:

• Achieved state-of-the-art accuracy in multi-class classification (up to 50.0% for one subject, 42.8% average).
• Deployed the VowelNet model on the BioGAP platform, achieving minimal power consumption (25.93 mW) and continuous operation for over 21 hours.
• Demonstrated classification latencies of 40.9 ms and identified temporal area electrodes as most crucial for accuracy.

Conclusions:

• This work represents a significant advancement toward practical, real-time, and unobtrusive speech imagery BCIs.
• The developed system enables new opportunities for covert communication and assistive technologies for individuals with speech impairments.
• The findings highlight the potential of ultra-low-power wearable devices and embedded machine learning for next-generation BCIs.