Thought to Text

Project Update – April 2026

The Thought to Text project is progressing well. Key developments this month:

• Data Collection App: Adapted Obadiah's data collection application for use by the Thought to Text team, streamlining our EEG recording workflow and enabling more consistent, structured data capture sessions.
• Improved AI Model: Integrated data filtering techniques to clean EEG signals prior to model input, significantly improving signal quality and reducing noise artifacts.
• EEGConformer Architecture: Upgraded the model architecture to EEGConformer — a transformer-based approach designed specifically for EEG classification — improving our ability to capture both local and global temporal patterns in brain signal data.

Project Overview – December 2025

The Thought to Text BCI project develops a brain-computer interface that directly translates EEG brain signals into text in real-time. Our goal is to help millions of individuals with severe motor disabilities — including those with ALS, cerebral palsy, and locked-in syndrome — who face significant communication barriers. Current assistive technologies average just 2–10 words per minute, far below the natural pace of human thought.

• Hardware: 16-channel OpenBCI Think Pulse system
• Architecture: EEGNet feature extraction + LSTM encoding
• Vocabulary: 30+ words for functional communication
• Performance target: Real-time inference under 100ms latency
• Signal flow: EEG Acquisition → Feature Extraction (EEGNet) → Temporal Encoding (LSTM) → Text Prediction

Modular Data Architecture

Our innovation centers on a unified inference engine operating across three distinct data sources — enabling rapid development and robust real-world deployment:

• ✓ Simulated Data (Complete): Synthetic EEG patterns with known ground-truth labels. Achieved 93–97% classification accuracy during initial development.
• ● Recorded Playback (Current Focus): Real EEG recordings played back at variable speeds (0.1x–10x), including labeled data for offline debugging.
• ⚙ Live Streaming (40% Complete): Real-time EEG acquisition with thread-safe buffered streaming, designed for sub-100ms continuous inference.

Key advantage: Train once on any data source — deploy everywhere without retraining.

Current Work: Recording Real Brain Signals

Playback System Operational:

• Recording pipeline with session management
• Variable-speed playback (0.1x – 10x)
• Label synchronization framework
• Offline evaluation and debugging tools
• Hardware-independent testing workflow

Hardware: OpenBCI Think Pulse 16-Channel Kit — wireless Bluetooth, 256Hz sampling rate, active electrodes for superior signal quality.

Currently recording team member EEG sessions to build a labeled dataset. Target: 100+ hours of diverse thought patterns capturing vocabulary variations and individual neural signatures.

Next Steps: Going Live (4–8 Weeks)

• Complete Live Data Pipeline: Finalize thread-safe streaming, implement real-time artifact rejection, and deploy adaptive filtering.
• Initial Live Testing Protocol: Conduct team member live inference sessions and benchmark end-to-end latency and classification accuracy.
• Data Collection Sprint: Record 50+ hours of labeled EEG across 5+ subjects, capturing diverse vocabulary patterns.
• Model Optimization: Fine-tune on real neural data, implement online learning, and develop subject-specific calibration.

Long-Term Vision (6 Months)

• Expand vocabulary to 50+ words for more comprehensive communication
• Develop user-facing GUI application for accessibility and ease of use
• Establish accessibility community testing partnerships
• Continue refining model performance with diverse user data
• Explore deployment pathways for individuals with communication disabilities