A silent word and a spoken word can begin with a related plan. A new study asks which parts of that plan are visible in brain activity. Reading it carefully means separating a laboratory classification result from the much harder task of helping someone communicate freely.
- A restricted task score needs its vocabulary, participants and recording conditions.
- Many recordings from one person do not create many independent participants.
- The next useful test would measure reliable communication, including mistakes and correction.
- Publication
- Nature Neuroscience · 29 September 2026
- Participants
- 9 people during awake brain-tumor surgery
- Recording
- High-density electrocorticography
- Task
- Imagined and overt syllables
Original publication: 29 Sep 2026 · The date above refers to this brief.
Syllable classification in the controlled task
Median accuracy, percent
What the study measured
Zhao and colleagues recorded cortical activity while participants imagined or articulated syllables. Some neural representations were shared between the tasks; others differed. The reported median classification accuracy was 80.4% for imagery and 78.3% for articulation. These are participant-level task summaries, not the percentage of ordinary thoughts a device can read.
The paper explicitly leaves broader vocabulary, continuous speech and long-term implanted use for further validation. Its controlled task is a research foundation, rather than a demonstrated everyday communication service.
Source 1 ↗Start by asking what the decoder is allowed to choose
A classifier choosing among a small set of prompted alternatives has a different job from recognizing an unprompted sentence. The menu, timing cue and training examples all help define the problem. A headline that leaves them out makes the result sound more general than the experiment.
When reading a future report, write down the available choices before the accuracy. Then ask whether the test includes pauses, unfamiliar words and moments when the person intends to say nothing. Those cases determine whether a system can distinguish communication from incidental activity.
A median is a middle result, not a guarantee
A median describes the middle of a set of values. It does not reveal the worst session, the most difficult syllable or the person for whom the system worked poorly. The bars here show reported medians without reproducing the original distributions. They should not be read as evidence that silent speech performs better.
For a practical application, a useful reporting set would include variation between people, variation within a person over time, error types and the effort needed to correct them. These are our evaluation questions, not additional outcomes measured in this article.
Connect decoding to the human workflow
Our chatbot coverage raises a useful parallel: an isolated model score and success with a person in the loop are different outcomes. A communication interface adds decisions about starting, stopping, rejecting an output and repairing a misunderstanding. Those choices deserve measurement alongside signal decoding.
A hypothetical device that produces a plausible phrase quickly but makes correction difficult could be less useful than a slower, more controllable one. That is an evaluation scenario, not a result from the present paper. The comparison table offers a way to keep the stages visible.
The next milestone worth watching
We would look for prospectively specified tests with new sessions, a larger task repertoire and clear rules for unsuccessful outputs. Participants should have an understandable way to control when the system acts. Reports should include failures and training burden rather than only selected audio demonstrations.
The realistic hope is increasingly dependable, user-controlled communication technology. The strongest next headline would describe an improvement in the complete communication task, with an appropriate comparison and honest uncertainty. A higher decoding percentage alone would leave much of that question open.
A syllable classifier and a communication system have different jobs
The current experiment, Figure 5 caption and data-access statement were checked. The older speech-to-text comparison uses the primary abstract and opening design description. The published scores are retained in their original units.
01What could participants choose to imagine?
- What was observed
- Five people used a six-syllable paradigm; four used an eight-syllable paradigm, during awake brain surgery.
- Where the conclusion stops
- A prompted menu is narrower than unrestricted thought or conversation. A single generic chance baseline would conceal the different menus.
Source 1 · Methods → Experiment paradigm
02Does 80.4% show superiority over overt speech?
- What was observed
- The reported medians are 80.4% for imagined and 78.3% for overt syllable classification; the paired comparison gives P=0.3008.
- Where the conclusion stops
- The 2.1-point difference between medians is descriptive. It is not the median paired improvement, a confidence interval or demonstrated superiority.
Source 1 · Abstract; Figure 5i caption
03Can another BCI score be placed on the same leaderboard?
- What was observed
- Willett et al. studied attempted speech-to-text in one participant with ALS and report word error rate, including a 125,000-word vocabulary.
- Where the conclusion stops
- Word error rate includes sequence errors and uses a different task. Do not convert it to syllable accuracy or pool these studies into one performance claim.
Source 2 · Abstract; Main, opening participant description
Numbers you can inspect
| Measure | Value & unit | Origin & method |
|---|---|---|
| Participants in the current study | 9 people | Reported Participants S1–S9; trial recordings remain nested within people. Source 1 · Methods → Experiment paradigm |
| Paradigm 1 participants | 5 people | Reported S1–S5 used six syllables. Source 1 · Methods → Experiment paradigm |
| Paradigm 2 participants | 4 people | Reported S6–S9 used eight syllables. Source 1 · Methods → Experiment paradigm |
| Imagined syllable classification | 80.4 median percent accuracy | Reported Participant median; uncertainty and individual distributions not extracted. Source 1 · Abstract; Figure 5i |
| Overt syllable classification | 78.3 median percent accuracy | Reported Same task family; not a separately randomized superiority trial. Source 1 · Abstract; Figure 5i |
| Difference between the two reported medians | 2.1 percentage points | Calculated 80.4 − 78.3 = 2.1. Difference of medians is not median of paired differences. Source 1 · Abstract; Figure 5i |
| Paired comparison in the current study | 0.3008 P value | Reported Two-sided Wilcoxon signed-rank test, imagined versus overt classification; not proof of equality. Source 1 · Figure 5i caption |
| Attempted speech, large vocabulary in the 2023 comparison | 23.8 percent word error rate | Reported 125,000-word vocabulary in one participant; not a syllable-classification error rate. Source 2 · Abstract |
Compare the actual experiments
These studies answer different questions. Read the unit and endpoint before comparing results.
| Study | Unit & setting | Readout | Interpretation boundary |
|---|---|---|---|
| Zhao et al. 2026 Source 1 · Experiment paradigm; Figure 5 | 9 people, awake intraoperative recording | Classification of prompted imagined/overt syllables | Restricted menus; the paper does not establish everyday independent communication. |
| Willett et al. 2023 Source 2 · Abstract; Main, opening design description | One participant with ALS; intracortical microelectrode arrays | Attempted speech decoded to text; word error rate | Different population, recording hardware, vocabulary and endpoint prevent a simple ranking. |
The export includes claims, available numbers, methods and source locations. It contains our reading notes and published summaries; it is not raw participant data or an independent reanalysis.
Evidence update · 2 Oct 2026
Added task-menu denominators, the exact paired-test result and an endpoint-preserving historical comparison. Raw patient data and the source-data workbook were not analysed. AI source check; no human editorial or clinical review.
Four checks for a speech interface
| What you see | What it can tell you | What to ask next |
|---|---|---|
| Prompted classification | Distinguishing a fixed set of alternatives | What happens with unfamiliar or absent input? |
| Repeated sessions | Stability within the same person | How much recalibration is required? |
| Communication task | Producing and correcting an intended message | Who controls start, stop and rejection? |
| Independent deployment test | Performance outside development sessions | Which errors and burdens remain? |
Our original reading framework. Questions and proposed checks are not reported experimental results.
Your questions, answered
Does this read private thoughts at a distance?
The reported experiment used cortical recordings during a structured task. It does not evaluate remote access to arbitrary thoughts.
Is 80.4% a patient success rate?
No. It is a median classification score for the study task, not a clinical outcome or a guarantee for every participant.
Why show two bars without declaring a winner?
A numerical difference alone does not establish superiority. The original distributions, paired comparison and intended question matter.
What would make a future result more useful?
A specified communication task with new sessions, clear error handling and measured user effort. A demonstration should include unsuccessful outputs as well as attractive examples.
Can I download the recordings behind these scores?
Our CSV contains published summary numbers and reading notes. It does not contain neural recordings. The primary paper places patient data under controlled access, with author and institutional review and a data-use agreement. Source-data tables and analysis code are linked by the paper, but we have not rerun them.
Limits of this interpretation
- Controlled syllable tasks do not establish open-ended or long-term communication performance.
- This account does not independently inspect the source-data distributions.
- Basic neural decoding research is not evidence that a particular clinical device is suitable for someone.
Sources & transparency
- Zhao, Wang, Liu et al. (2026): A neural architecture for imagined and overt speech motor dynamics
Publisher abstract, Experiment paradigm methods, Figure 5 caption, decoding results and data/code availability checked. The two paradigms use six and eight syllables. Patient data are controlled access; statistical source-data workbooks and code were not reanalysed. Reported medians are newly drawn, not a publisher figure. · Accessed 2 Oct 2026
DOI: 10.1038/s41593-026-02456-0 - Willett, Kunz, Fan et al. (2023): A high-performance speech neuroprosthesis
Primary abstract and opening participant/microelectrode description checked. Used only to contrast task, participant population and endpoint; no clinical device appraisal, transcript reanalysis or supplement assessment. · Accessed 2 Oct 2026
DOI: 10.1038/s41586-023-06377-x
Prepared and source-checked with AI; source access recorded on 2026-09-30 (UTC). Press-news Team is our collective publication byline, not a medical reviewer. No human editorial or clinical review has taken place. We did not conduct these experiments or reanalyse their raw data. Reported findings, our explanations and proposed follow-up tests are distinguished. Access limits appear with each source. This is an educational account of basic research and research methods, not an individual diagnosis or treatment recommendation. Photographs are illustrative. Evidence notes added by AI from the access scope stated below. Calculated values use the printed arithmetic; published point estimates are not an independent reanalysis. No human editorial or clinical review.
Source check: AI source check — linked primary passages, denominator notes and deterministic arithmetic
Clinical review: Not applicable to this educational guide
Suggest a correction