--- title: "China built a speech brain–computer interface that decodes Mandarin for locked-in patients" date: 2026-09-28 category: Brain–Computer Interface site: NeuroAI canonical: https://neuroai.site/a/na-bci-als-speech-decoding language: en --- # China built a speech brain–computer interface that decodes Mandarin for locked-in patients > Two separate Chinese teams have pushed speech brain–computer interfaces (脑机接口) past a wall that English-first research never had to climb — decoding tonal Mandarin in real time. One system already let a 67-year-old ALS patient rebuild everyday phrases; another spelled out Chinese characters from neural signals at 49.6 characters a minute. A 67-year-old woman who had lost the ability to speak to ALS sat in a Beijing hospital and thought about water. A fraction of a second later, the words "我要喝水" — "I want water" — appeared on a screen. No voice, no keyboard, just a thin film of electrodes resting on the outside of her brain. For decades, speech brain–computer interfaces (脑机接口) were a story told in English. The hard part was always assumed to be the algorithm. In China, researchers hit a different wall first: Mandarin is tonal, monosyllabic, and dense with homophones, so decoding it is not a translation of an English system — it is a different problem. Two projects in 2025 showed that problem is now partly solved. ## The patient who got her words back In March 2025, a team led by Zhao Guoguang (赵国光), director of Xuanwu Hospital (宣武医院) and the National Center for Neurological Disorders, implanted a semi-invasive system called **Beinao-1 (北脑一号)** into the left language-motor region of that patient's brain. The device uses a **128-channel flexible electrode** laid on the dura mater — outside the brain tissue, not inside it — with a coin-sized control and transmitter embedded in the skull. The results, published through Xinhua and Beijing municipal science channels: - **Single-character decode latency under 100 milliseconds** — fast enough to feel like talking, not typing. - After just **three hours of training**, the system decoded 62 common words at 34% accuracy. It has since reached **52% real-time accuracy** and can produce phrases like "我要吃饭" ("I want to eat") and "今天心情很好，我想和家人散步" ("I'm in a good mood today, I want to walk with my family"). - The first participant was a **67-year-old ALS (amyotrophic lateral sclerosis / 渐冻症) patient** with severe anarthria — exactly the locked-in population these systems are built for. By May 2026, Beinao-1 had been implanted in **16 people** across several Beijing hospitals, accumulating **more than 55,000 hours** of safe operation, with ALS patients decoding close to **100 Chinese words**, according to the official Beijing science-commission briefing. ## The system that spelled Mandarin at speed The deeper scientific claim came from Shanghai. A group led by neurosurgeon **Wu Jinsong (吴劲松)** at Fudan University's Huashan Hospital (华山医院), engineer **Zhou Zhitao (周志涛)** at the Shanghai Institute of Microsystem and Information Technology (中科院上海微系统与信息技术研究所), and **Tao Hu (陶虎)** of the BCI company BrainTreasure (脑虎科技) published in *Science Advances* on 5 November 2025 the first real-time, full-spectrum decoding of Mandarin. Their setup used a **256-channel flexible electrocorticography (ECoG) array** — far more channels than Beinao-1, placed directly on the cortical surface during a clinical procedure. Key numbers: - **394 of the roughly 400 Mandarin syllables** covered, decoded purely from neural signals. - **71.2% median syllable-identification accuracy** in single-character reading. - **65-millisecond latency** per syllable. - **49.6 characters per minute** real-time sentence decoding — about one-fifth of normal speech speed, but genuine live output, not post-hoc analysis. *Science* itself covered the paper, noting the work "opens [BCIs] up to huge populations of potential patients" because most of the world's languages are tonal. ## Why Mandarin is the hard mode of speech BCI English speech decoders lean on phonemes — around 40 to 50 of them. Mandarin packs meaning into **syllables plus pitch**: the same syllable spoken with a different tone means a different word. A decoder that ignores pitch produces gibberish. The Shanghai team solved this with a dual-stream architecture that processes syllable and tone in parallel, then layers a trigram language model on top. That matters beyond China. Hundreds of millions of people speak tonal languages — not only Mandarin but Cantonese, Thai, Vietnamese, and most sub-Saharan African languages. An English-first decoder leaves them out. A Mandarin decoder, once it works, is the template for all of them. ## What is still missing The Shanghai decoding trial ran on **one participant — a 43-year-old woman with epilepsy whose speech was normal**, not a locked-in patient. She read prompts aloud; the system keyed off the sound onset. It did **not** decode silent, imagined speech, which is what a profoundly paralyzed person actually needs. Beinao-1's ALS result is real but limited to a small vocabulary at ~52% accuracy. So the honest picture is two halves of one tool: Shanghai proved the *language* can be decoded in real time; Beijing proved the *clinical* path works in an actual ALS patient, at small scale. Neither yet delivers fluent, silent, open-vocabulary speech. ## What readers can do now - **If you work in assistive tech or clinical neurology**, treat Chinese speech-BCI papers as a fast-moving frontier, not a curiosity. The *Science Advances* decoder and the Beinao-1 implant are the two reference points to track. - **If you build for tonal-language markets**, watch this space: a Mandarin decoder is the hardest tonal case solved first, which means other tonal languages are next. - **If you report on BCIs**, separate the "decoded from a healthy speaker" result from the "helped a locked-in patient" result — conflating them is the most common error in this story. ## Honest limitations Core facts verified across authoritative sources: the Beinao-1 (北脑一号) semi-invasive 128-channel implant and its 67-year-old ALS patient, single-character latency <100 ms, 52% real-time accuracy, and phrases decoded are from Xinhua (20 and 24 March 2025) and the Beijing Municipal Science & Technology Commission (25 March 2025); the 16 implants / 55,000+ hours / ~100 words figure is from *Science and Technology Daily* (11 May 2026). The Shanghai real-time Mandarin decoding — 256-channel ECoG, 394 syllables, 71.2% median accuracy, 65 ms latency, 49.6 char/min, published 5 November 2025 in *Science Advances* (DOI 10.1126/sciadv.adz9968) — is confirmed by the paper abstract, CAS Shanghai, *Science*, *Medical Xpress*, and *People's Daily*. Crucial caveat: the *Science Advances* decoding subject was a 43-year-old epilepsy patient with normal speech who read aloud; it was not an ALS or locked-in participant, and silent/imagined speech was not decoded. Beinao-1's ALS result uses a small closed vocabulary at ~52% accuracy. No RMB amounts appear in this article, so no currency conversion applies. This analysis is current to 28 September 2026 and is not medical advice. --- Published by NeuroAI (https://neuroai.site/) — https://neuroai.site/a/na-bci-als-speech-decoding Free to quote with attribution and a link to the original.