Imagined Speech Detection Using Multi-Receptive CNN for Asynchronous BCI Communication and Neurorehabilitation

Imagined speech-based brain-computer interface (BCI) facilitates brain signal-driven intuitive communication which holds great promise as an effective speech rehabilitation tool, enabling real-time, hands-free interaction for individuals with speech and motor impairments. While speech-based assistant systems rely on wake-word detection (e.g., "Hey Siri"), BCI-based communication system must capture imagined onset from EEG signals to turn on the 'brain switch' to further convey user's imagined command. Nevertheless, the absence of reliable ground truth for the endogenous paradigm adds to the complexity to train the model to capture exact onset from continuous EEG. To address these issues, we introduce a multi-receptive field convolutional neural network, designed to capture speech and idle states based on behaviorally-aligned EEG features. We propose a voice-based ground truth alignment method with voting strategy that aims to synchronize imagined speech with overt speech onset and offset, providing a structured approach for capturing speech events in asynchronous BCI systems. Furthermore, spectral and phonological analyses revealed that beta and alpha bands, as well as syllable count, appear to influence speech state discriminability. Evaluations on imagined and overt speech tasks, including pseudo-online experiments, demonstrate the potential to enhance asynchronous BCI systems, supporting real-time communication for both healthy and impaired individuals.