BCIwiki (bciwiki.com) — The study was published in Computers in biology and medicine on June 18, 2026. The researchers propose dynamic wavelet-basis selection that minimizes wavelet entropy for each EEG epoch, selects a more informative basis, and injects Gaussian noise into the corresponding coefficients. A convolutional neural network with channel-wise excitation then classifies the augmented imagined-speech signals.
The dataset contains 32 channels, 8 stimuli, and recordings from 10 participants. The words-vowels combination achieved a highest classification accuracy of 98% with a Cohen's kappa of 0.95, while performance was lower for the full class set. The researchers report that dynamic wavelet augmentation outperformed conventional augmentation strategies and static wavelet approaches, helping the model handle noise and variation in EEG. The result provides another method option for data augmentation and silent-speech decoding in non-invasive imagined-speech brain-computer interfaces.