Comparison of single-channel EEG decoding performance in imagined speech tasks

Electroencephalography (EEG)-based imagined speech decoding technology offers a communication solution for patients with language impairments. However, existing research on English-language systems cannot be directly applied to Chinese due to significant linguistic differences. Multi-channel EEG systems face hardware complexity and computational delays. This study aimed to develop a single-channel EEG decoding system for Chinese imagined speech, explore neural mechanisms to promote low-cost rehabilitation devices.

EEG signals were recorded from four Chinese participants imagining five distinct Chinese characters. The SGLCNN model was proposed and compared with VGG16, EEGNet, and machine learning methods. Experiments evaluated the performance of 20 single-channel time–frequency graphs and compared their accuracy with traditional methods.

The CP5 and C3 channels achieved the highest classification accuracies of 74.33% and 73%. The SGLCNN model outperformed EEGNet and VGG16 in decoding binary classification tasks for Chinese characters. Single-channel signals improved classification accuracy by 0.67% (SGLCNN) and 2.75% (EEGNet) compared to full-brain signals under the same models.

This is the first study to decode imagined Chinese speech using single-channel EEG and tests the performance of 20 channels for this task. The SGLCNN system maintains performance while reducing hardware needs. Identifying dominant channels (CP5/C3) informs clinical electrode placement, accelerating brain-computer interfaces (BCIs) translation.