Electrode Selection and Convolutional Attention Network for Recognition of Silently Spoken Words from EEG Signals

Brain signals generated during silent speech have shown to be useful in designing a communication-based brain computer interface (BCI). However, brain signals are non-stationery and complex in nature, and therefore challenging to recognize. We propose a framework for recognizing imagined words using brain signals captured through electroencephalograph (EEG) sensors. Our method consists of two main components: (i) an electrode selection method, and (ii) a convolutional attention network. The electrode selection method provides the electrodes containing the most discriminative time-frequency information for imagined speech recognition. Further, spectrograms from selected electrodes are used as input to the convolutional attention network, that extracts time-frequency features and performs classification by ascribing higher importance to the time points with higher discriminatory capacity. Experimental results using EEG dataset shows that the proposed method efficiently recognizes mentally spoken words and exhibits performance superior to that of state-of-the-art methods.