Inspiration
Every year, millions of people lose their ability to speak due to ALS, Parkinson’s, or other neurodegenerative diseases. These individuals remain fully conscious. Aware of everything around them yet unable to communicate even the simplest needs like "yes", "no", "water", "food", or even "help". We wanted to give them part of their voice back, to help them regain independence and connection using brain computer interface technology.
That’s how NeuroAlpha was born, a system that can translate thoughts into words in real time.
What it does
NeuroAlpha is a working prototype designed to decode imagined speech, the words a person is thinking, into text. This differs from attempted speech, where a person mentally tries to speak or move speech muscles without producing sound; imagined speech involves no articulatory movement at all. Using EEG brainwave data from the Neuropawn headset, our model detects neural activity patterns associated with specific words and predicts what the user is thinking in real time.
How we built it
We started from scratch since no imagined speech dataset existed; we collected our own EEG data using the Neuropawn headset. Dry spike electrodes were placed on the frontal lobe, and participants imagined four key words - Yes, No, Water, and Food. We gathered 300+ samples, including background noise, at 125 Hz. We then trained an LSTM model with residual blocks to learn temporal EEG patterns. After tuning dropout and optimizing the loss function, we achieved about 70% accuracy. Finally, we built a real-time interface that visualizes brain activity and predicts the imagined word live on screen.
Challenges we ran into
- Limited datasets: While a few small imagined-speech EEG datasets exist, most are research-grade and collected using wet-electrode or high-density systems. Their formats and electrode placements weren’t compatible with our affordable Neuropawn hardware, so we had to design and record our own dataset from scratch.
- Signal noise: EEG data is extremely sensitive. We had to reduce interference using bias fibers and electrode adjustments.
- Hardware limitations: Dry electrodes need good scalp contact; hair density affected signal quality.
- Subject variability: Each person’s brain patterns differ, so individual calibration was necessary.
Accomplishments that we're proud of
- Built the first working imagined-speech-to-text model with affordable hardware.
- Achieved ~70 % accuracy distinguishing between three classes
- Took the first step toward giving people with speech impairments a non-invasive way to communicate.
What we learned
- EEG data is highly individual, meaning large-scale data will be key for a generalized decoder.
- Words that activate the same parts of the brain would be hard to classify. The frontal lobe placement was effective but adding more electrodes could improve this problem.
- And most importantly, we learned that it is possible to read imagined speech. This isn’t science fiction anymore!
Next Steps
- Expand dataset & vocabulary: Collect more EEG data across diverse users to improve generalization and enable decoding of a larger vocabulary of words and phrases.
- Refine neural decoding models: Experiment with advanced architectures such as transformer-based temporal models and attention-driven EEG encoders for higher accuracy in imagined speech recognition.
Log in or sign up for Devpost to join the conversation.