VowelNet: Enhancing Communication with Wearable EEG-Based Vowel Imagery

For decades, researchers have explored Speech Imagery - silently imagining speech - as a communication aid for those with severe impairments. Despite advancements in classification accuracy, existing methods mainly rely on offline, resource-intensive machine learning techniques that necessitate multiple channels, leading to obtrusive setups and social stigma, preventing their application outside clinical settings. This paper presents, for the first time, vowel imagery classification based on a low channel-count, ultra-low-power wearable EEG system (BioGAP), and VowelNet, a novel lightweight neural network optimized for real-time Speech Imagery processing (up to 6 classes) on compact, low-power System-on-Chips. VowelNet requires 8x fewer channels than current EEG-based speech recognition systems, and it provides accuracies up to 91.1% for vowel-rest classification and up to 61.8% for inter-vowel classification. When running on the edge (GAP9), it enables continuous speech imagery classification for more than 1 day on a small 150 mAh LiPo battery, with an output latency of only 41 ms. This work paves the way for non-stigmatizing and energy-efficient assistive communication devices.