Improving EEG Signal Interpretation for Phoneme Classification through Residual Convolutional Neural Network

Phoneme classification in speech imagery Brain-Computer Interfaces aims to decode imagined speech signals directly from brain activity, enabling communication without overt speech. This challenging task requires accurately distinguishing between various phonemes from complex, non-stationary EEG data, offering potential advancements in assistive communication technologies for individuals with speech impairments. This paper presents an approach for classifying phonemes in imagined speech tasks using EEG signals by employing an advanced deep learning model. The proposed methodology centers around a convolutional neural network architecture enhanced with residual connections, designed to improve classification performance by acquiring complex temporal and spatial patterns related to the EEG signals. Unlike conventional machine learning techniques, the residual connections enables the model to train deeper networks effectively, mitigating performance degradation even with a limited dataset. This approach is implemented using a speech imagery EEG dataset, achieving a marked improvement in classification accuracy over traditional methods. The results underscore the potential of this model for advancing multiclass phoneme classification in EEG-based brain-computer interface systems for future work towards imagined speech recognition.