ИИ декодирует смысл речи из активности мозга и переносится на новых людей без дообучения

In July 2026, researchers presented on arXiv a framework for cross-subject semantic decoding that reconstructs the meaning of perceived speech from invasive brain recordings and transfers to a new person without a single retraining step.

What exactly the scientists did

The team trained a system to translate neural activity into contextual semantic embeddings — the same vector representations of meaning used by language models. The data were collected using electrocorticography (ECoG): electrodes are placed directly on the surface of the cortex while a person listens to and understands natural speech.

The key technique is a shared latent space. Brain responses from multiple participants are mapped into one shared space using a shared response model (SRM), and the decoder learns to predict semantic embeddings from the aligned signals rather than from the "raw" signals of a specific person. This shift matters: the system predicts not literal words but vectors of meaning, so it relies on responses to speech content that are shared across people, rather than on the individual characteristics of the signal.

• Data: invasive ECoG recordings during comprehension of natural speech
• Alignment method: shared response model (shared latent space)
• Decoder target: contextual semantic embeddings, not individual words
• For a new participant: individual projection + pretrained decoder without retraining
• Result: consistently above baseline methods, smaller accuracy drop on an "unfamiliar" brain

Why generalization across people is hard

The main problem of invasive neural decoding is that brains are literally not comparable: each person has their own electrode configuration, their own anatomy, and their own pattern of neural signals. A model trained on one participant usually loses accuracy sharply on another.

The authors work around this as follows: for a held-out participant (who was not part of training), only their individual projection into a predefined shared space is estimated, while the decoder itself is used as-is. According to the paper, this noticeably narrows the quality gap between a "familiar" and an "unfamiliar" brain compared to baseline approaches, and reduces the accuracy drop when moving from the original participant to a new one.

"Aligning neural activity in a shared latent space when decoding into

the space of semantic embeddings proves to be an effective strategy for improving cross-subject generalization," the paper's abstract on arXiv states.

What this means

The approach outlines a path toward neural interfaces that don't need to be calibrated for weeks for every new patient: the shared space and the semantic decoder transfer over, and only a thin projection is adapted to the person. For medical speech prostheses, this potentially means a faster start-up and less data per user — though for now this is a research prototype, not a clinical device.

Frequently asked questions

What is cross-subject decoding?

It is the reconstruction of information from brain activity so that the model works not only on the person it was trained on, but also on new people. In this work, the goal is to decode the meaning of perceived speech from invasive ECoG recordings.

Does the model need to be retrained for a new person?

No. For a new participant, the authors estimate only their individual projection into the shared latent space and apply the already pretrained decoder without retraining.