Brain-to-Language Decoding: Tasks, Signals, Methods, Evaluation, Practical Use and Beyond
Yiqian Yang, Yiqun Duan, Chenyu Liu, Yiqi Wang, Xinliang Zhou, Chin-Teng Lin, Yu Zhang
cs.CL · cs.HC · cs.NE
Quick verdict
Rigorous survey linking brain-to-language decoding tasks, signals, methods and evaluation; exemplary critical stance on attribution, but no new empirical data.
Summary
This survey critically synthesises brain-to-language decoding — recovering text, speech and facial animation from neural activity during articulated, inner and perceived language tasks — across invasive and non-invasive recordings. Its method is a PRISMA-S-informed critical narrative synthesis: Europe PMC and OpenAlex searches (28,515 records; no lower year limit; cut-off 15 September 2026) produced a 920-entry registry, 246 entries with structured extractions. Content is organised by task condition (three groups, nine subtypes), recording family (EEG through intracortical, fMRI, fNIRS/HD-DOT, fUSI) and three decoding routes mapping neural features to phonemic units, acoustic/articulatory targets or semantic embeddings. Within matched protocols it compares clinical systems (intracortical 125,000-word decoding at 23.8% WER; home use exceeding 3,800 hours and 1.96M words at 99.2% cued-copy accuracy) with non-invasive retrieval (41% top-1 in the best MEG retrieval setting) and text generation, showing teacher-forced reports (ZuCo BLEU-4 up to 18.89) inflate against autonomous re-evaluations (WER above 108%); semantic-decoder margins over noise controls can be as small as BERTScore 0.830 versus 0.819. Limitations: no new empirical data or meta-analysis; only 246 of 920 entries source-checked; four commercial databases unsearchable; many 2025-2026 citations are preprints; the five-level future trajectory is speculative.
Review complete
Assessment at a glance
4
risk flags identified
● high 0
● medium 2
● low 2
5 strengths · 5 weaknesses · 5 author questions
AI text signal = Binoculars zero-shot detector score over 20 sampled passages (Fast-DetectGPT co-reported) — absence does not prove human authorship, and a flag does not imply misconduct.
Levels reflect verification complexity — not a quality score. Evolve never rates your paper.
Key claims under review
- Task conditions (Articulated, Inner, Perceived), not model architecture alone, determine what linguistic information a decoder can learn, so comparisons must be made at the condition level.
- Phonetic, acoustic and semantic decoding targets preserve different aspects of a message and are complementary routes rather than competing maturity stages.
- Many published non-invasive text-decoding gains depend on test-time teacher forcing, familiar stimulus text or leaky splits; under autonomous generation with proper controls, word error rates exceed 100% (e.g., COFETT reruns in Figure 14d).
- Online communication depends on calibration, feedback and user control in addition to decoding accuracy, as shown by longitudinal home use (over 3,800 hours and 1.96M words at 99.2% cued-copy accuracy, Card et al. 2026).
- The field's future can be organised as a five-level trajectory from commands and language towards meaning, scenarios and bidirectional cognitive exchange.
Strengths
Operational task taxonomy with explicit boundary rules
Table 2 and Figure 4 define Articulated/Inner/Perceived with nine subtypes and adjudication rules: attempted speech without confirmed residual articulation is neither Articulated nor Inner; a written cue does not make a later interval Reading. This prevents the task-label conflation that plagues most prior surveys.
Documented, auditable review protocol
§2.3 and Appendix C report 12 Europe PMC plus 4 OpenAlex query groups, pagination to exhaustion, 28,515 database records and 20,335 citation-tracked records, yielding a 920-entry registry with per-entry provenance labels (Figure 3d); failed Scopus/Web of Science/IEEE Xplore/ACM DL access is disclosed rather than hidden.
Attribution and leakage discipline in evaluation
Matched-protocol comparison of clinical systems
Honest boundary-setting for emerging modalities
Weaknesses
No new empirical or meta-analytic contribution
Uneven source verification across the corpus
Retrieval gaps in engineering venues
Heavy reliance on unreviewed 2025-2026 preprints
Conflict-of-interest status not visible in extracted text
Full Review Report
$1.99 · one-timeUnlock the complete preliminary analysis Evolve produced for this paper.
Report preview
Scope & Claims
This is a critical narrative survey of brain-to-language decoding — recovering text, audible speech and facial animation from neural activity during Articulated (Overt/Mouthed/Limited), Inner (Articulatory/Auditory/Mixed) and Perceived (Auditory/Visual/Audiovisual) tasks. §2.1 excludes EMG-only and lip-reading interfaces, spelling-by-selection BCIs, and auxiliary state detection. Three claims anchor the synthesis: (1) task conditions, not architectures, determine what a decoder can learn; (2) phonetic, acoustic and semantic targets preserve complementary aspects of a message; (3) online communication depends on calibration, feedback and user control alongside accuracy. A prospective five-level trajectory (§9) extends from commands to bidirectional cognitive exchange.
The remaining sections cover Methodology Assessment, Statistical Evidence, Results & Reproducibility, Limitations & Risks, and Verdict & Recommendations.
- All 5 strengths & 5 weaknesses, analyzed in detail
- 4 risk flags with severity, section & reviewer questions
- Methods & statistical evidence assessment
- 5 questions a peer reviewer would ask
- Full six-dimension evaluation report & recommendations
One-time purchase · this paper only
One-time purchase, this paper only. AI-generated preliminary analysis — not peer review.
Frequently asked questions
What is brain-to-language decoding?
It is the translation of neural activity recorded during speaking, imagining speech, or listening and reading into outputs such as text, synthesised speech or facial animation. Clinically it aims to restore communication for people who have lost speech; scientifically it probes how the brain represents language.
How accurate are implanted speech brain-computer interfaces?
Reported clinical systems decode vocabularies from about 50,000 to 125,000 words at word error rates of roughly 2.5-25% with output rates around 60-80 words per minute; one long-term home user produced 1.96 million words over more than 3,800 hours with 99.2% accuracy on cued copy sentences. Figures are protocol- and participant-specific, mostly from single-user studies.
Can EEG alone generate text from brain activity?
Not reliably as autonomous generation. Word and segment retrieval from EEG/MEG is above chance (about 41% top-1 in the best MEG settings), but when reading EEG-to-text systems are re-evaluated without teacher forcing — where models are fed correct previous words during testing — word error rates exceed 100%, meaning outputs are largely wrong relative to references.
Is this paper a systematic review?
No. It is a PRISMA-S-informed critical narrative synthesis with a documented search (28,515 records retrieved; a 920-entry registry) but no preregistration, no dual independent screening and no exhaustive full-text review. Several commercial databases (Scopus, Web of Science, IEEE Xplore, ACM DL) could not be searched.
About this review
This is an AI-generated preliminary review by Evolve.Pub, produced from the openly available text of the paper. It is not peer review and does not represent the authors' views. The canonical record of the paper is at the publisher link above.
- Extracted text is truncated at multiple points; Sections 6-9 (evaluation, practical use, directions, trajectory), parts of §2.4 and §5.5-5.6, and several tables (e.g., Tables 11, 15, 19-21) were unavailable — this review concentrates on §§1-5.6.
- Figures are assessable only via captions; schematic content and heatmap values could not be independently verified.
- Many cited works are 2025-2026 preprints whose peer-review status and final numbers could not be verified here.
- Analysis-code availability is not stated in the extracted text; only the data workbook and registry exports are described as accompanying artifacts.
Go deeper on this paper
Wrote a paper? Stress-test it before reviewers do.
Upload your manuscript for the same structured scrutiny — reasoning, statistics, evidence, figures, and reproducibility.