Abstract
This study investigates the decoding and comparison of brain signals associated with spoken Thai and English words using deep learning techniques and EEG equipment. In the field of Brain-Computer Interfaces (BCI), researchers have extensively explored methods to decode brain signals into text. Two primary approaches exist: invasive (e.g., ECoG) and non-invasive (e.g., EEG). Invasive methods require surgery and offer highquality signals but carry infection risks. Conversely, non-invasive methods employ scalp electrodes, resulting in lower signal quality but greater practicality for daily use. The present research utilizes three datasets each for Thai and English to evaluate the effectiveness of EEG and compare the outcomes for both languages. The Thai word data consists of three sets: single words (หิว, ปวด, เจ็บ, หนาว, ร้อน), two-word phrases (หิวมาก, ปวดท้อง, เจ็บแขน, หนาวมาก, ร้อนมาก), and three-word sentences (ฉันหิวมาก, ฉันปวดท้อง, ฉันเจ็บแขน, ฉันหนาวมาก, ฉัน ร้อนมาก). The English word datasets correspond semantically to each Thai set. All results are tested and compared using two machine learning approaches: Multi-Layer Perceptron (MLP) with statistical features and Convolutional Neural Network (CNN) with stacked spectrogram features. The MLP achieved an overall accuracy of 98%, while the CNN achieved 64%.
Authors
Adsawinnawanawa, E., Faragy, S., Buasod, S., & Keeratipranon, N.
DOI:10.4186/ej.2025.29.4.53