A 45-year-old man with ALS sat in a lab and tried to speak. No sound came out — his disease had taken his voice years earlier. But on a screen, words appeared within milliseconds, and a synthesised voice, cloned from recordings made before he fell ill, spoke them aloud with his own pitch and rhythm, even managing a sung melody. The "reader" was not a person. It was a large model (大模型) trained to translate patterns of brain activity into language.
This is the year brain-computer interface (脑机接口) stopped being only about moving cursors and started being about meaning.
The Western breakthrough: speech you can hear, and thoughts you can hide
Two strands of 2025 research defined the field. At the University of California, Davis, a team led by Maitreyee Wairagkar and Sergey Stavisky implanted 256 electrodes in the brain region that coordinates speech for a patient with ALS. Reported in Nature in June 2025, the system reached about 97% accuracy on a 50-word vocabulary after just 30 minutes of training and synthesised speech within roughly 25 milliseconds — fast enough for real-time conversation. The voice was reconstructed from the patient's pre-illness recordings, and he could modulate tone to ask questions or sing.
At Stanford, researchers pushed past attempted speech into inner speech — the silent words we think but never say. Published in Cell in August 2025, the study worked with four participants who had severe speech impairment and decoded imagined sentences with up to 74% accuracy on controlled tasks, and a word-error rate as low as 26% on a 125,000-word real-time vocabulary. Crucially, the team built a "thought password": users had to imagine a private rare phrase before decoding switched on, blocking accidental reads of inner thoughts with about 98.75% reliability. At UCSF, Edward Chang's group went further, generating not just voice but an animated digital avatar that reproduced facial expression and emotion.
The throughline is machine learning. Brain signals are noisy and personal; neural networks learn each user's unique patterns and stitch phonemes into words in real time.
China's answer: a "Chinese operating system" for the brain
Chinese teams entered 2025 with a distinct challenge. English has roughly 40 phonemes; Mandarin's combinations of initials, finals and four tones push the count past 400, making decoding harder. Two independent efforts tackled it head-on.
In July 2025, the Shanghai Yansi Brain-like AI Research Institute (上海岩思类脑人工智能研究院) together with Huashan Hospital of Fudan University (复旦大学附属华山医院), the National Center for Neurological Disorders, Zhejiang University and the Shanghai Institute of Microsystem and Information Technology of the Chinese Academy of Sciences reported an implantable EEG large model (植入式脑电大模型). With 8–10 stereotactic EEG electrodes, a patient needed only about 100 minutes of training on 54 Chinese characters to produce fluent sentences. The model breaks speech into initials and finals and recombines them, reaching expression across nearly 2,000 common characters, parsing a full sentence within half a second, with reported accuracy above 83% for initials and 84% for finals and a 1:36 extrapolation ratio.
Separately, the Shanghai Institute of Microsystem (CAS), Huashan Hospital and NeuroXess (脑虎科技) achieved real-time Chinese decoding published in Science Advances in November 2025: a 256-channel flexible brain-computer interface (脑机接口) reached 71.2% accuracy across 394 syllables, with 65 ms latency and a real-time rate of 49.6 characters per minute. Science itself highlighted the work as a step toward letting speakers of tonal languages talk again after stroke or disease.
According to Li An, Chief Scientist at BrainNet (脑机网), China's authoritative AI observatory, the Chinese focus on tonal-language decoding is less a follower's move than a deliberate bet on the hardest linguistic case, which doubles as a strong baseline for other languages.
Why it matters to ordinary readers
For the estimated millions who lose speech to stroke, ALS or brain tumours, this is the difference between isolation and conversation. But the same capability that restores voice also raises the privacy questions covered in our neural-data piece: a model that can read intended speech can, in principle, read unspoken thought. The Stanford "thought password" is the field's first practical acknowledgement that mental privacy must be engineered in, not bolted on.
Honest limitations
The headline accuracies come from small participant groups — often a single patient or a handful — and from staged tasks, not free daily conversation over months. Chinese results were demonstrated in epilepsy patients undergoing clinical procedures, not in long-term implant users with degenerative disease. The "large model" framing describes neural-network decoders trained on brain data; they are not general chatbots. Real-time inner-speech decoding still degrades on open-ended, spontaneous thought, and no system yet reliably reads unfiltered imagination. All figures are research-reported and await larger, independent replication.
What readers can do now
- Families of stroke or ALS patients should ask neurologists about speech brain-computer interface (脑机接口) trials rather than assuming the technology is unavailable.
- Anyone using consumer EEG should understand that today's consumer headsets decode attention or sleep, not language — claims of "mind reading" are almost always marketing.
- Watch for the thought-password-style safeguards as a minimum standard before trusting any speech-decoding product.
- Follow Nature, Science and Cell papers (and their Chinese-institution counterparts) for verified milestones rather than vendor demos.