NEST

Non-invasive brain-to-text · Open source

From EEG to open-vocabulary text

NEST is a hardware-agnostic decoding stack that turns non-invasive EEG recorded during reading into text, and reports its numbers the way the field should: free decoding, held-out subjects, noise controls.

Research prototype. No validated benchmark yet. The first word error rate ships together with its evaluation protocol.

nest decoder
Input (EEG, 105 channels, 8 frequency bands)
Output (text)

press start to decode the illustrative sample

Subject: held-out
Features: 840 / word
Decoder: BART, free decoding
WER: pending
Illustrative: synthetic waveform, ZuCo reference sentence, not model output.

Status · September 2026

Where the project stands

Every number on this site is verified against the repository. When something is not measured yet, it says pending.

24,137
lines of Python across preprocessing, models, training and evaluation
12,071
ZuCo sentence-subject pairs processed by the data pipeline (66 GB raw corpus, 53 files, 12 subjects, 105 channels)
184M
parameters in NEST v2 BART (64.7M trainable); v2 CTC 45.3M
pending
word error rate. The first full v2 run is an 8 to 12 hour A100 job that has not been run yet

We withdrew a number. In February 2026 we published a word error rate that had been estimated from training loss, not measured by decoding. It was removed. Results appear here only when they are decoded on held-out subjects with a noise-input control. Read the evaluation protocol.

Why now

Non-invasive brain-to-text left the lab in 2025

Three things changed at once: the research results, the capital, and the shape of the gap.

01

The results moved

Meta's Brain2Qwerty decoded typed sentences from MEG at a 32% character error rate in February 2025. The June 2026 v2 release reached 61% word accuracy. Those are MEG numbers on a closed, typing-style task; the same study's EEG condition scored 67% character error rate. Non-invasive decoding is improving quickly, and open-vocabulary EEG reading, the task NEST works on, is still unsolved.

02

The capital arrived

Venture funding for brain-computer interfaces went from roughly $0.4B in 2024 to about $1B in 2025. In the first seven months of 2026 the non-invasive share rose to 29%.

03

The gap is a software layer

Implant companies operate on patients. Headset companies sell hardware. The one vertically integrated text decoder, Sabi, only runs on its own cap. Nobody ships an open decoding layer that works on any EEG system.

How it works

Word-aligned EEG in, sentence out

NEST follows the ZuCo feature convention: for every word a reader fixates, 105 channels times 8 frequency bands become one 840-dimensional vector. A sequence of those vectors is one sentence. ZuCo is by Hollenstein et al. (2018), released under CC BY 4.0.

01
EEG recording
105-channel scalp EEG while the subject reads naturally.
(channels=105, 500 Hz)
02
Word-aligned features
Eye-tracking fixations align EEG to words; 8 bands per channel.
(words, 840)
03
Transformer encoder
Projection, positional encoding, 6 pre-norm layers.
d_model=768, heads=8
04
BART decoder
Cross-attention over the EEG sequence; language prior from pretraining.
facebook/bart-base
05
Text
Free autoregressive decoding. No teacher forcing at test time. Greedy today, beam search planned.
no teacher forcing

Hardware-agnostic

NEST is software. It takes preprocessed features, so any research-grade EEG system that can produce them is a valid input. Transfer to low-channel consumer headsets is an open question we have not tested.

Subject-independent

Train on 8 subjects, validate on 1, test on 2 the model never saw. Personal calibration is a later feature, not a hidden assumption.

Open

MIT-licensed code, data loaders and evaluation scripts. The benchmark protocol is part of the product.

Evaluation

The field's numbers were inflated. Ours are built to survive that critique.

EEG-to-text papers on ZuCo reported BLEU-1 around 40. In 2024 the community found those scores came from teacher forcing, and that random noise scored about the same as real EEG.

Published

Teacher-forced scores

Wang and Ji (AAAI 2022) BLEU-1 40.1. DeWave (NeurIPS 2023) BLEU-1 41.4. BELT 42.3. All measured with the decoder given the correct previous word at every step.

Corrected

Free-decoding scores

Jo et al. 2024 showed noise input matches real EEG under that protocol. A July 2026 replication reports free-decoding WER above 99% for the same model family. Open-vocabulary EEG-to-text is unsolved.

What NEST commits to.

  • Free autoregressive decoding at evaluation time, never teacher forcing. The code path is public; today it is greedy, beam search is planned.
  • Held-out subjects that never appear in training or validation.
  • A noise-input control reported next to every real-EEG number.
  • The protocol published together with the first result, as a technical report.

What exists

Built in seven months, verified in the repository

ComponentTodayStatus
Codebase24,137 lines of Python: band-pass and ICA preprocessing, six model architectures, training and evaluation scriptsImplemented
DataZuCo pipeline written and run on the real corpus: 66 GB raw (53 .mat files), 12 subjects, 105 channels, 12,071 sentence-subject pairs, NaN handling for unfixated words. The raw data is re-downloaded before the GPU runImplemented
ModelsNEST v2 CTC (45.3M parameters) and NEST v2 BART (184M total, 64.7M trainable) instantiate and runVerified
EvaluationFree autoregressive decoding (greedy today), greedy CTC decoding, word- and character-level edit distanceImplemented
Trainingv1 model trained 100 epochs on 12,071 real samples (5.4 h on an M2 laptop); loss decreased, no WER computed. v2 needs a GPUPending GPU
DeploymentPruning, quantization, ONNX and TorchScript export, streaming inference modules writtenUnbenchmarked
Tests80 test functions; suite currently broken by code drift and is being repairedIn repair

Who it is for

Research tool first, clinical product later

Sequenced the way non-invasive neurotech companies that reached revenue actually did it.

Phase 1

EEG research labs

A decoding SDK and an open benchmark for language-EEG research. Outside medical-device scope by design: no diagnostic or therapeutic claim.

Phase 2

Clinics and corpora

An ethics-approved Turkish reading corpus and research-only pilots with rehabilitation and neurology departments. The first Turkish EEG-to-text dataset.

Phase 3

Assistive communication

A CE-marked communication aid for people with ALS or locked-in syndrome who keep reading comprehension, delivered with EEG hardware partners.

Roadmap

The next twelve months

Each quarter ends with a measurable artifact: a benchmark report, a corpus release, a license count, a pilot protocol.

Q4 2026
Validated benchmarkGPU training of v2; free-decoding WER with noise control; arXiv technical report.
Q1 2027
Turkish corpusEthics approval; data collection with 20 participants on a prepared protocol; SDK v0.1.
Q2 2027
CompanyIncorporation in the Teknopark Istanbul Cube incubator; TUBITAK BiGG application; first lab licenses.
Q3 2027
PilotResearch pilot with a rehabilitation center; first EEG hardware integration.

FAQ

Straight answers

Does it work today?
Not yet in the sense that matters. The architecture, data pipeline and evaluation code are built and verified. The v2 model has not been trained to completion because that needs a GPU we do not have. No word error rate is published until it is measured with free decoding on held-out subjects.
Reading is not the same as intending to speak. Do you know?
Yes. ZuCo records EEG while people read sentences shown on a screen, so a model trained on it decodes responses to perceived text. Decoding a sentence a person is trying to produce is a different, harder problem that needs inner-speech data. Phase 2 of the roadmap targets exactly that, and until then we do not claim to decode intent.
Is NEST a medical device?
No. NEST is a research tool and makes no diagnostic or therapeutic claim, which keeps it outside FDA, EU MDR and Turkish TITCK device scope. A future assistive communication product would go through CE marking under MDR, and we plan for that in phase 3, not before.
Which EEG hardware does it need?
Any research-grade system that can deliver word-aligned frequency-band features. The reference data is ZuCo, recorded with a 128-electrode system of which 105 channels are used, plus eye tracking for word alignment. NEST does not make hardware.
How is this different from Neuralink or Synchron?
They implant electrodes surgically and reach far higher accuracy because of it. NEST works from the scalp. The trade is accessibility for signal quality, and the honest position is that non-invasive open-vocabulary decoding is still an open research problem.
Why Turkish?
Every public EEG-to-text corpus is English. A Turkish natural-reading corpus collected under ethics approval at Istanbul Commerce University would be the first of its kind and a defensible asset for the company.
What about the brain data of participants?
EEG is health data. Any recording we make in Turkey needs ethics approval, explicit KVKK consent for a stated purpose, the special-category security measures the Turkish data protection board requires, and storage inside Turkey unless a lawful transfer mechanism exists. The public ZuCo corpus is de-identified and CC BY 4.0 licensed.
Why open source?
Because the field's credibility problem is a measurement problem. Public code, public protocol and public controls are the only way a claim in this space can be checked.

Follow the first benchmark

Watch the repository for the v2 training run and the technical report.