Deep Learning Networks for Vowel Speech Imagery

Speech Imagery (SI) is a successful alternative for communication systems based on Electroencephalographic (EEG) signals that do not need external stimuli like evoked potentials. A recent strategy for SI is to analyze Speech Related Potentials (SRP) features on EEG signals to recognize vowels. However, there is research to be done in the development of machine learning methods for vowel classification in SI signals based on SRP. Therefore, this paper proposes two neural networks to classify vowels in speech imagery signals using SRP: a Convolutional Neural Network called sCNN and a Capsule Neural Network called sCapsNet. The experiments were developed with the DaSalla dataset. According to the results, sCapsNet reports better performance than sCNN and Support Vector Machine (SVM). The average accuracy was 71.9%, 67.63%, and 71.33% respectively. Besides, the capsules of sCapsNet could model by vectors the SRP features regardless of time and differences of SI vowels by subjects.