NeuroAI NEUROAINEUROAI.SITE
China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories
ESC

The surgeon keeps the patient awake, then a computer reads the syllables

A Huashan Hospital neurosurgeon spent years mapping language in awake patients, yielding a 256-channel electrode array that decoded Mandarin syllables mid-surgery.

2026-11-16 · 1145 words · NeuroAI
The surgeon keeps the patient awake, then a computer reads the syllables

In a typical awake craniotomy, the patient counts aloud. A small electric current touches a patch of cortex for two seconds. The counting stops — not because she is hurt, but because the words are suddenly not there. The current is switched off, the counting resumes, and the surgeon marks that patch of tissue as a place not to cut.

Neurosurgeons have done this for decades, most often during tumour surgery near the language areas; glioma (胶质瘤) is the classic case. It is the only reliable way to tell which millimetres of cortex a patient cannot afford to lose. The procedure is called an awake craniotomy (清醒开颅手术).

Jinsong Wu (吴劲松), a professor of neurosurgery at Huashan Hospital, Fudan University (复旦大学附属华山医院), has spent much of his career inside that procedure. In late 2025 he became the senior author of a paper that turned the same setup into something else: a brain-computer interface (脑机接口) that decodes Mandarin in real time.

The operating room as a laboratory

An awake craniotomy gives researchers something no lab can manufacture: a human brain, open, with electrodes on the cortex, and a patient who can talk. The window lasts an hour or two. Everything has to be ready before the first incision.

That constraint shaped the collaboration. Wu's team worked with engineers at NeuroXess (脑虎科技), the Shanghai company co-founded by Tiger H. Tao (陶虎), which builds flexible electrodes that sit on the cortex surface rather than penetrating it. The electrodes used in the published work were a 256-channel microelectrocorticography (µECoG) array — a sheet of sensors laid on the surface of the brain, not spikes pushed into it.

The surgical logic stayed the same as always. The electrode goes where the clinical procedure already needed to go; the research rides along.

What the 2025 paper showed

"Real-time decoding of full-spectrum Chinese using brain-computer interface" appeared in Science Advances on 5 November 2025 (vol. 11, no. 45, eadz9968) and is open access. The headline numbers, taken from the abstract:

  • 256 channels of surface cortical recording.
  • 394 distinct Mandarin syllables decoded purely from neural signals.
  • 71.2% median accuracy for syllable identification in a single-character reading task.
  • Real-time sentence decoding demonstrated on top of that syllable decoder.

The participant was one patient. The paper describes its method as a "tonally integrated, direct syllable" decoding approach — tone is part of the unit being recognised, rather than a property inferred afterwards.

The ambition behind it is written into the paper's first sentence: speech BCIs are aimed at people with anarthria — patients who cannot speak because of ALS or a brainstem stroke — not at healthy volunteers wearing headsets.

Why Mandarin needed its own decoder

Most speech-decoding research is in English and works on phonemes. English has roughly forty of them, and a decoder that can string phonemes together gets a long way.

Mandarin does not decompose that way. It is tonal and monosyllabic: a small set of syllable units, each carrying a tone that changes meaning, and the mapping from sound to character is many-to-one. Decoders designed for phoneme sequences do not transfer cleanly, and the published Chinese work has historically been much thinner than the English literature.

That is why the paper's unit choice matters more than its accuracy figure. If you decode syllables directly and handle tone as part of the unit, you are building the layer that a Chinese speech prosthesis actually needs — rather than a system that must be rebuilt later.

What 71.2% means in practice

Seventy-one percent is not a product. It means roughly three syllables in ten are wrong at the unit level, before any language model cleans up the output. A sentence assembled from that stream needs redundancy, context, and probably a large model (大模型) sitting downstream to make it readable.

That is exactly the architecture the field is converging on: a neural decoder that is imperfect, plus a language model that is very good at guessing what a person probably meant. Whether that division of labour is acceptable clinically is an open question, because the model's corrections can also quietly replace what the patient actually wanted to say.

The other constraint is context. The published sentence decoding was a laboratory protocol with fixed reading tasks, which is a different thing from following a conversation that the decoder has never seen before.

The gap between a paper and an implant

The clinical path is long and it is now regulated in the open. China's medical-device rules treat implantable BCIs as high-risk devices, and any commercial system needs a registration trial with multiple sites, durability data, and long-term safety follow-up.

Wu's contribution sits at the clinical end of that pipeline. He is a surgeon who can place an electrode and a researcher who can design a decoding experiment around a patient who is already on the table. That combination is scarcer than either skill alone, and it is why the published Chinese results tend to come from a small number of hospitals rather than from companies working alone.

There is also a quieter point about who these systems are for. The patient in this study, like most participants in the published Chinese literature, needed brain surgery anyway — in this case an epilepsy operation in a patient whose tumour sat near the language area. The scientific gain is real; so is the ethical obligation to be clear that a research electrode placed during a necessary operation is not the same as an implant offered as a treatment.

Honest limitations

Everything quantitative here comes from one peer-reviewed paper with a single participant, and single-participant decoding results are notoriously fragile across people, electrode placements, and day-to-day signal stability. The 71.2% figure is a median on a constrained reading task, not a conversation.

I did not interview Wu or any patient. The description of his role as senior author is from the paper's author list; his departmental titles are as listed in a 2023 seminar biography published by Tsinghua's McGovern Institute, and institutional titles in Chinese hospitals change often enough that a reader should treat them as dated.

I also left out claims I could not verify. Company announcements about later trials, faster decoding speeds, and regulatory progress circulate widely but were not, at the time of writing, backed by peer-reviewed papers I could read, so they are not repeated here.

What readers can do now

  • Read the paper itself. Science Advances made it open access (DOI 10.1126/sciadv.adz9968); the methods section explains what a 256-channel µECoG array can and cannot record.
  • Separate demonstrations from devices. When you see a BCI headline, look for three things: how many participants, whether the task was constrained, and whether the result is peer-reviewed.
  • If you are a patient or caregiver, ask about trial registration and inclusion criteria rather than about demo videos; the clinicaltrials.gov and Chinese registry entries are where eligibility is actually defined.

Related coverage

Cite this article

NeuroAI (脑机网), "The surgeon keeps the patient awake, then a computer reads the syllables," 16 November 2026, https://neuroai.site/a/na-people-wu-jinsong-mandarin-bci

Machine-readable: BibTeX · RIS · JSON · Markdown

More in “Builders” → · Back to home · Markdown version

China AI, reported clearly
1,300+ verified stories on models, robotics, BCI and the business of AI.
Browse stories