Millions of people with medical conditions such as paralysis, Amyotrophic Lateral Sclerosis (ALS), also known as Lou Gehrig’s disease, or severe speech impairments are unable to easily communicate with the world around them. Even simple tasks like typing or speaking can become impossible, leaving them dependent on slow or limited assistive technologies. A project by researchers Ovishake Sen, Raghav Soni, Darpan Virmani, and Baibhab Chatterjee, PhD, aims to change that by enabling communication directly from brain activity.
The work develops a system that can read brain signals and convert them into written characters. A participant wears a non-invasive electroencephalogram (EEG) cap and imagines writing letters. The EEG signals are cleaned, preprocessed, and converted into meaningful features, which are then classified by a lightweight neural network model.
A user can image letters, for example, H – E – L – P, which can then be transcribed onto a screen. This technology has the potential to be highly impactful for individuals with paralysis, ALS, locked-in syndrome, or severe speech and motor impairments by allowing them to communicate without relying on speech or mobility.
This work shows that people could one day ‘write with their thoughts,’ enabling faster, more natural communication for individuals who cannot speak or type, and bringing brain-computer interface technology closer to everyday use.”
– Ovishake Sen, PhD student researcher
A key contribution of this work is that the entire system runs in real time on a small, portable edge device (NVIDIA Jetson TX2), instead of requiring a powerful computer or cloud server. This makes the technology more practical, faster, and accessible for real-world use. The proposed system achieved about 89.83% accuracy using 85 extracted EEG features, and a smaller 10-feature version reduced latency to about 202.62 ms per character with less than 1% accuracy loss.
The technology enables the development of practical, portable brain-computer interface (BCI) systems that utilize EEG technology without requiring invasive implants or large computing infrastructure. This approach paves the way for future hands-free human-computer interaction interfaces, allowing users to engage with computers or devices through neural intent without the need for physical typing or speech.
This project won first place in the student Demo Competition at the Warren B. Nelms Annual IoT Conference in December 2025. The team received very positive feedback, especially around the fact that the system is not only an offline EEG classification model, but a real-time edge-deployable system. Attendees were particularly interested in how the project balances accuracy, latency, and portability, and they appreciated the use of non-invasive EEG instead of implanted electrodes, which makes the approach safer and more scalable. There was strong interest in the real-time deployment on the NVIDIA Jetson TX2 and the ability to reduce latency using feature selection.
A portion of this project has been published in Scientific Reports (Nature Publishing Group), highlighting the significance and impact of this work in the field of brain–computer interfaces and real-time neural decoding.
In the future, the research team hopes to expand the participant pool and collect a larger, more diverse EEG dataset. This would help to improve cross-subject generalization so the model works better for new users with minimal recalibration. They’d also like to move from single-character prediction toward more continuous text generation and phrase-level communication. Discussions at the conference helped reinforce that cross-subject robustness, larger participant studies, and wearable integration are the most important next steps.
This work demonstrates one of the first real-time, portable edge-device implementations for imagined handwriting recognition from non-invasive EEG signals. The result is important because it moves BCI research beyond offline model evaluation and toward deployable systems that can operate with realistic latency and power constraints.
Reference
Sen, O., Soni, R., Virmani, D., Parekh, A., Lehman, P., Jena, S., Katikhaneni, A., Khalifa, A., & Chatterjee, B. A low-latency neural inference framework for real-time handwriting recognition from EEG signals on an edge device. Scientific Reports, 15, 41040 (2025). https://doi.org/10.1038/s41598-025-24972-y