Recognition of Bengali Vowels from Auditory Evoked Potentials Using CNN

Audio evoked potentials, or AEPs, have recently gained popularity. It can be described as the EEG signal generated when a person is exposed to auditory stimuli. The human brain behaves differently when exposed to diverse sounds. It also produces a distinct hemodynamic activity pattern for speech sounds compared to non-speech sounds. This paper presents a system for capturing AEP. The proposed approach covers seven different Bengali vowels. A total of 208 sets of data were obtained from five different subjects. The various AEP signals were then filtered and pre-processed. Different features Power Spectral Density (PSD), Entropy, and Variance have been extracted from the preprocessed signal. Convolutional Neural Network (CNN) has the highest accuracy (90.96 %) for PSD features. The dataset has been made public for future research purposes.