Get Started
ABOUT THE PROJECT

From Brain to Text

NEST uses a transformer-based encoder-decoder architecture to decode EEG signals into natural language. Here is a complete overview of our pipeline, model, and research context.

MISSION

Why We Built NEST

Brain-computer interfaces hold transformative potential for people with motor or speech disabilities, for scientific research into cognition, and for the future of human-computer interaction.

Existing approaches require invasive surgery, proprietary hardware, or closed-source implementations. NEST is our answer: a fully open, non-invasive, reproducible framework that researchers and engineers can build on.

By publishing weights, datasets, and complete training code under an MIT license, we aim to accelerate the entire field of EEG-based BCI research.

Open Research
Full codebase, checkpoints, and training scripts publicly available
Reproducible Results
Deterministic evaluation with fixed random seeds and released test splits
Benchmark Standard
Evaluated on ZuCo — the most comprehensive EEG reading dataset available
Community Driven
MIT licensed, contributions welcome from the global research community
PIPELINE

End-to-End Architecture

NEST transforms raw EEG recordings into natural language through a four-stage pipeline designed for accuracy and speed.

01
EEG Input

105-channel EEG at 500Hz. Captured during natural reading with standard research-grade headset.

02
Preprocessing

Bandpass filtering (0.5–100Hz), artifact removal via ICA, epoch extraction aligned to word onsets.

03
Transformer

6-layer EEG encoder with 8-head attention. Cross-attention decoder generates word-level text tokens.

04
Text Output

Natural language text decoded from brain signals, ready for NLP downstream tasks.

MODEL

Architecture Details

The NEST model is built on a standard transformer architecture adapted for the unique structure of EEG time-series data.

EEG Encoder

6 Transformer Layers
512 hidden dimensions, 8 multi-head attention layers. Processes temporal and spatial EEG patterns into a rich latent representation.

Cross-Attention Bridge

8 Attention Heads
Learns alignment between EEG feature sequences and text token positions. Enables the decoder to selectively attend to relevant brain patterns.

Text Decoder

6 Transformer Layers
50,000 vocabulary size. Autoregressive generation with beam search decoding and length normalization.

TRAINING

Training Details

Trained on the ZuCo dataset with 12 subjects, 400+ natural reading sentences, and 100 epochs of supervised learning.

12,071
Training Samples
100
Epochs
5.4h
Training Time
32
Batch Size
ZuCo Dataset
12 subjects reading natural English text (newspapers, Wikipedia) while 105-channel EEG is recorded at 500Hz. The gold standard benchmark for EEG-to-text research.
REQUIREMENTS

Hardware & Software

NEST is designed to run on standard research hardware. No exotic setups required.

Minimum Requirements
  • Python 3.9+
  • PyTorch 2.0+
  • 8GB RAM
  • CPU-only inference (slow)
  • 10GB disk space
Recommended for Training
  • NVIDIA GPU (RTX 3080+)
  • CUDA 11.8+
  • 16GB+ GPU VRAM
  • 32GB System RAM
  • 50GB disk space
READY TO EXPLORE?

Start Your BCI Research

Dive into the demo, read the research paper, or clone the repo and start training your own models.