The purpose of this project was to develop a neural network algorithm capable of recognizing subvocalized (thought to oneself) letters.

Individuals with Amyotrophic Lateral Sclerosis (ALS) and locked-in syndrome encounter enormous challenges in communicating with others, significantly limiting their independence and quality of life. Brain-computer interface (BCI), a technology that decodes electrical information from the brain, has recently been used in some experiments to restore communication in individuals unable to speak or write. BCI often requires invasive and expensive surgical implants of neuroprostheses.

Brain electrical data were collected from a male healthy individual thinking a letter, either A, B, or C, to himself (subvocalization) using a commercially available, relatively inexpensive, non-invasive EEG. A total of 298 samples were collected, uploaded to a Python notebook, and augmented. The first 238 samples (first 80%) of each augmented array were set off for training, the rest of the samples were discarded except for the 60 samples in the original array for testing (none from the augmented ones). Each of these training arrays were then concatenated into one array with length 238×30 = 7140 samples. A 1-dimensional convolutional neural network (CNN) model was used to classify each sample based on the label.

The CNN model with the best performance had 63.33%, classifying 38/60 samples correctly with a statistically significant p value of 1.9316 × 10⁻⁶. For the subvocalized letter “A,” 13/20 samples were correctly identified, for “B,” 12/20, and for “C,” 13/20.

Results from this pilot study show that a CNN model can be used to learn from data of subvocalized letters collected using a relatively low-cost EEG with statistically significant accuracy. Future research should consider additional data collection from multiple individuals including ALS patients, use of a medical-grade EEG, and better CNN hyperparameters to achieve higher accuracies.