Study decodes open-vocabulary words from silent reading using 49 hours of dry EEG

An arXiv paper that reached the Hacker News front page reports that word-level information can be decoded from brain recordings made with a dry-electrode EEG cap while a person silently reads. Submitted to arXiv on 20 August by Ingo Marquardt, the study rests on roughly 49 hours of recordings from a single participant, and its central claim is twofold: open-vocabulary decoding of silent reading works with non-invasive hardware, and the main thing holding performance back is training-data volume, not an absence of signal.

Why inner speech is hard to collect

The paper begins with a data problem that has stalled inner-speech decoding. As the authors point out, there is no way to assemble a corpus that pairs neural activity with someone's spontaneous inner monologue, because no one can verify what a person was actually thinking at a given moment. The usual substitutes, asking people to repeat a cued word in their head or having them generate inner speech and report it afterwards, are slow to gather, weakly time-locked, and compliance cannot be checked. The authors instead treat silent reading as a stand-in: the presented text is known in advance, the timing is fully controlled, and the cognitive act is close enough to inner speech to serve as a practical proxy.

The experiment

According to the paper, one participant completed 393 sessions totalling around 49 hours of 19-channel dry-electrode EEG, covering approximately 240,000 word presentations. The words came from continuous narrative text, shown one at a time in rapid serial visual presentation. The typography of each word was randomised on every trial, a design choice meant to partially separate a word's identity from its low-level visual appearance and make it harder for a model to succeed by recognising letter shapes rather than processing the word itself.

The model

The decoding pipeline combines a convolutional encoder over short EEG windows with an optional causal transformer, trained with a CLIP-style contrastive objective. Rather than predicting words directly, the system learns to line up each EEG segment with the hidden-state embedding that a large language model produces for the word being read. Performance was measured as top-10 retrieval evaluated per word group against shuffled-label baselines, in effect asking whether the true word lands among the model's ten best candidates more often than chance would allow.

What the results show

Decoding was reliably above chance and extended to mid-frequency and rare words, not just the most common vocabulary. Performance rose in proportion to the logarithm of the amount of training data, with no plateau in sight, which underpins the authors' argument that the approach is constrained by dataset size rather than a hard ceiling in the neural signal.

Two secondary findings stand out. Removing the occipital and posterior-temporal electrodes, which sit over visual processing areas, cut the word-level decoding gain by about a third while leaving the model's ability to track narrative context intact. And a set of control analyses was needed to disentangle genuine word-level decoding from two potential confounds: the model's ability to follow the story's broader context, and a non-neural positional bias introduced by the transformer's positional embeddings.

Caveats worth noting

The headline results come from a single participant, so there is no evidence yet that the same pipeline transfers across people, and treating silent reading as a stand-in for inner speech remains an assumption rather than a demonstrated equivalence. Top-10 retrieval is also a long way from reconstructing sentences, since the model narrows down candidates rather than identifying exact words. Finally, 49 hours of dense sampling is far beyond what most EEG studies collect, which cuts both ways: it demonstrates the value of scale, but it suggests the approach currently demands an unusual amount of recording from each user.

Why it matters

High-bandwidth speech brain-computer interfaces have so far relied on electrodes implanted in or on the brain. According to this paper, a cheap 19-channel dry cap can carry recoverable, open-vocabulary lexical signal, provided there is enough data and the right contrastive training against language-model embeddings. For assistive communication, where surgery is a major barrier to adoption, that is a meaningful datapoint, even if decoding someone's unprompted thoughts remains a far harder and more distant problem than decoding what they happen to be reading.