AI training system provides promising advance in decoding imagined speech from brain electrical activity
Posted on 17 June 2026
A new system for training a deep learning AI model to ‘read’ imagined speech from EEG recordings has improved the accuracy of the technology, new research published in Neuroscience shows.
Neuroscientists and computer scientists around the world are trying to find a way to interpret the words someone is thinking of purely by analyzing brain electrical activity using EEG signals. The goal is to provide patients who have lost the ability to speak because of a stroke or another condition that causes paralysis affecting speech with a means to communicate again.
But reading imagined speech – where someone imagines a word or letter without saying it out loud – from EEG signals is not straightforward. One the biggest challenges inter-subject variability. When one person thinks of a word or letter, the EEG signal can look very different to when someone else thinks of the same thing. This means that current systems for reading imagined speech have limited accuracy and require time-consuming calibration to individuals.
New training approach
In the study published in Neuroscience, researchers at Kutahya Dumlupinar University in Turkey tested an artificial neural network they had developed, NeuroSilentia, for its ability to decode imagined speech from EEG signals, along with a new system for training the model aimed at addressing inter-subject EEG variability. This new AI training system forces machine learning models to identify patterns of EEG signals characteristic of specific imagined letters and seen across different people, rather than homing in on EEG patterns specific to one individual.
To test their system, Safa Dörterler a Research Assistant at the university who led the research, along with fellow computer scientists Emrullah Şahin and Durmuş Özdemir, used an imagined speech dataset released in 2024 in which 30 volunteers were asked to think of 31 Arabic characters while EEG signals were recorded using a headset with 14 electrodes. This data set, with the large number of characters, was more complex and therefore realistic data than that used in many similar studies in which volunteers thought of fewer letters.
When the researchers first tested NeuroSilentia, which weights EEG channels in data for how important they are in predicting a letter someone is thinking of, its accuracy was 20%. This was an improvement over other EEG models tested on the same data, which were only seven to 15% accurate.
Enhanced accuracy
“The main improvement came from our training strategy, Dynamic Subject-Invariant Fragment Mixing, or DSIFM,” says Dörterler. In DSIFM, segments of EEG readings from the same letter recorded in different individuals are split into short fragments during model training and grouped together. “We try to stop the model from asking ‘which person does this signal come from?’ and instead force it to focus more on ‘which imagined letter does this EEG signal represent?’”
When DSIFM was used to train NeuroSilentia, the model was 51% accurate in predicting which letter someone was thinking of. With cleaned data, removing the trials with the noisiest EEG readouts, the accuracy increased further – to 60%.
“Some people may ask whether this accuracy is sufficient for real-world use. The honest answer is that this is not yet a practical clinical communication system. However, the task is a highly challenging 31-class imagined-speech EEG problem with low signal quality and strong subject variability. In that context, the improvement should be interpreted as a meaningful methodological step rather than a finished application.”
The research published in Neuroscience is part of the TÜBİTAK project, led by Associate Professor Özdemir, which is developing a thought-to-text system using EEG signal processing, AI models and language processing methods, all based on Turkish EEG datasets. “This paper contributes by focusing on one of the core technical barriers – how to make EEG imagined speech models less dependent on the individual person,” says Dörterler.
Important next steps
“It is very important not to present this as ‘mind reading’ or as a technology that is already solved. We are still at an early stage. EEG-based thought-to-text is an exciting and promising idea, but current systems are not yet reliable enough for everyday clinical use. We need better data, stronger models, more robust generalization and much more validation,” says Dörterler.
“DSIFM needs to be tested on more datasets, different languages and different EEG devices. In this study, we focused on a 31-class Arabic alphabet imagined-speech EEG dataset. Future work should examine whether the same strategy works for Turkish, English or other languages, and for letters, digits, words and sentence-level imagined speech tasks.
“This is one of the directions we are pursuing in our own project. We are collecting EEG data with Emotiv devices with different channel numbers, including 5-, 14- and 32-channel systems. This will allow us to compare low-density wearable EEG settings with higher-density recordings and better understand how robust these methods are under different recording conditions.”
Dörterler studied computer science for his bachelors degree, MSc and PhD, researching AI, swarm intelligence and metaheuristics. “I have increasingly combined this background with EEG signal analysis, brain-computer interfaces and imagined speech classification,” he says. “This interdisciplinary area is exciting because it is not only about improving an algorithm, it is also about trying to understand complex human brain signals and, in the long term, contributing to technologies that may directly improve people’s lives.”
Dörterler says the researchers chose to publish the study in Neuroscience because it is at the intersection of AI, EEG, computational neuroscience and brain-computer interfaces. “We wanted the work to reach not only engineering and computer science readers, but also a broader neuroscience audience interested in EEG, cognitive neuroscience, computational neuroscience and brain-computer interface research. Neuroscience was therefore a suitable journal for the study.”
This article was written by Dr. Andy Ridgway.
About Neuroscience
Established in 1976, Neuroscience is the flagship journal of IBRO. The journal features papers describing the results of original research on any aspect of the scientific study of the nervous system. Papers of any length are considered for publication provided that they report significant, new, and carefully confirmed findings with full experimental details.
Together with IBRO Neuroscience Reports, IBRO’s open access journal, Neuroscience plays a crucial role in supporting the organization’s global neuroscience activities, as proceeds from both journals support more than 90% of IBRO’s initiatives.
As Neuroscience celebrates 50 years of advancing brain research, the journal is also marking the milestone with a global campaign highlighting and celebrating neuroscience around the world – explore the anniversary campaign to learn more.
