NEST

Demo

An illustrative decoder, labelled as such

This page shows what NEST's input and output look like: a synthetic EEG waveform on the left, a sentence typed out on the right the way a decoder would produce one. It does not run a trained model and it is not a benchmark result.

This page runs no model. The waveform is synthetic and the sentences are ZuCo reference sentences. It exists to show the shape of the input and output, not a result.

EEG input
eeg signal
Input (EEG, 105 channels, 8 frequency bands)
Channels
105
Bands
8
Features per word
840
Sampling
500 Hz
Alignment
eye-tracking fixations
Decoded text
nest decoder
Output (text)

press start to decode the illustrative sample

Synthetic waveform, ZuCo reference sentence, not model output.
Decoder
BART, free decoding
Subject
held-out
WER
pending
Latency
not measured

Evaluation protocol

What a real result will look like

When NEST v2 is trained and evaluated, results appear here in this shape: free decoding on held-out subjects, next to a noise-input control, under the same metrics.

ConditionWERCERBLEU-1ROUGE-1
Real EEG, held-out subjectspendingpendingpendingpending
Noise input controlpendingpendingpendingpending
Differencependingpendingpendingpending

A single row is not evidence. In 2024 the field found that a decoder can score well on real EEG and just as well on random noise pushed through the same pipeline, because a pretrained language model can produce fluent text from almost any input. NEST reports real EEG and a noise-input control side by side, under the same free-decoding protocol. Only the difference row, real EEG against noise, says anything about whether the model is reading a brain signal rather than guessing from its language prior.

Status

Why there is no live model yet

NEST v2 has no trained checkpoint. The only completed training run used an older, smaller architecture, about 2.78 million parameters, for 100 epochs on the full ZuCo set; its loss decreased but no word error rate was computed. The v2 CTC and BART models, 45.3 million and 184 million parameters, instantiate and run correctly, but training either to convergence needs a GPU run estimated at 8 to 12 hours on an A100, which has not happened yet.

Once a v2 checkpoint exists, this page will not need a server to demonstrate it. The repository already includes an ONNX export module, part of the evaluation code, that can convert a trained checkpoint for inference in the browser. Until that checkpoint exists, this page stays illustrative: a synthetic waveform and a scripted sentence, clearly labelled as such.

See the code behind this page

The training run, the evaluation protocol and the export module referenced above are public.