Neural Noise

Decoding individual words from non-invasive brain recordings for speech prosthesis

2 min read

  • Word-level decoding from non-invasive brain recordings (e.g. EEG) is advancing toward speech and communication prosthetics.1
  • Such approaches are highly relevant for locked-in and speech-restoration applications and support credible regulatory pathways for assistive devices.1
  • Nature-reported work on decoding individual words from non-invasive recordings is tier-1 for EEG-based BCI and speech prosthesis.1 1

Weekly enrichment (2026-07-20)

  • The primary work (d’Ascoli et al., Meta AI/FAIR, Nature Communications, 2025) introduces a deep-learning pipeline that decodes individual words from electro- (EEG) and magneto-encephalography (MEG) signals.23
  • The model was evaluated on nine datasets (seven public plus two collected by the authors), totaling 723 participants reading or listening to roughly five million words across three languages.2
  • It reached up to 37% top-10 accuracy on a fixed 250-word retrieval set and could decode words absent from the training set (zero-shot), performing significantly above chance (p < 0.005).2
  • MEG and reading were consistently easier to decode than EEG and listening, and accuracy scaled with the amount of training data and with averaging multiple responses at test time, indicating signal-to-noise is the main bottleneck.2
  • Benchmark context: with a 50-word vocabulary the non-invasive decoder reached 20% top-1 accuracy, versus 39.5% reported for an intracranial motor-cortex electrode, quantifying the invasive vs non-invasive performance gap.2
  • The reduced 250-word vocabularies covered between 70% and 95% of all word occurrences across datasets, contextualizing task difficulty.2
  • A key limitation: the study decodes language perception, not production, so it is not yet a direct speech-prosthesis solution for people who cannot speak.2
  • BCI implication: a single architecture generalizing across devices, languages, and tasks is a concrete step toward scalable non-invasive speech BCIs, though single-trial performance still lags intracranial neuroprostheses.24

Footnotes

  1. https://news.google.com/rss/articles/CBMiX0FVX3lxTE1sLW9oSnZ1dF9pVGxFTmlEeF92SG9reng3U0lYd2hJOVFZQkhrTlRzV0VVczRwd2VPMEFfVFl5WTRrUWJPeVI5X05xbjRHRGxxZ2UtQVM2SkVjYmxfaktB?oc=5 ↩ ↩2 ↩3 ↩4

  2. https://www.nature.com/articles/s41467-025-65499-0 ↩ ↩2 ↩3 ↩4 ↩5 ↩6 ↩7 ↩8

  3. https://arxiv.org/abs/2412.17829 ↩

  4. https://ai.meta.com/blog/brain-ai-research-human-communication/ ↩

Graph View