BRESLAU, 1874 — A German neuroanatomist named Carl Wernicke sat with a group of patients who had one thing in common: a wound in the same strip of brain, just above and behind the left ear. They could hear perfectly well. They could also speak — fluently, at a normal pace, with the rhythm of ordinary conversation.
The problem was what came out. Words arrived in the wrong places. A patient meaning to say “shoe” said “foot.” Someone reaching for “leg” produced “peg.” Sentences dissolved into what clinicians still call, without much charity, word salad. Wernicke wrote up the cases that year in a short monograph, Der aphasische Symptomenkomplex, and proposed something that would sit in textbooks for a century and a half: that this patch of temporal cortex holds the “auditory image of words” — a template of how a word should sound, from which its meaning and its pronunciation could both be recovered.
Then the field simplified him. The patch became Wernicke’s area, the listening centre. Speech production was handed to Broca’s area, in the frontal lobe, and the two were drawn as separate rooms joined by a corridor of nerve fibres. The picture was clean enough to survive 150 years of neuroscience. On 19 November 2026, in Science, a team at the University of California, San Francisco published the experiment that finally breaks it — and the answer was hiding in a space about the size of a nickel.
The paradox Wernicke himself could not explain is embarrassingly simple. If this region is a listening device, why does destroying it also destroy speaking? Every neurology student learns the bedside signs: comprehension gone, repetition broken, output fluent but wrong.
Clinicians filed the production deficit under comprehension — the patient speaks badly, the reasoning went, because he cannot hear himself and correct his own errors. That is the “proofreader” hypothesis, and it is testable: if Wernicke’s area only monitors speech after the fact, its activity during speech should follow the sound, not precede it. For decades nobody could settle the timing.
Functional MRI resolves brain activity over one to two seconds; the events in question happen in a few hundred milliseconds, and the difference between planning a word and hearing yourself say it is smaller still. “The basic question has been, ‘Why would damage to a hearing area of the brain break the ability to speak?’” the study’s senior author, the neurosurgeon Edward Chang, said at a briefing ahead of publication. “That question has been open for 150 years.”
Chang’s laboratory did not set out to overturn a textbook. It set out to use one. For fifteen years the lab has worked with epilepsy patients who come to UCSF for surgery and spend part of the operation awake, with a grid of electrodes laid directly on the exposed surface of the brain — high-density electrocorticography, the same clinical procedure surgeons use to map which tissue can be cut without costing the patient language.
The electrodes record neural populations at a resolution an MRI cannot approach, millisecond by millisecond. To that the team added a newer instrument: Neuropixels probes, recording the activity of single neurons. Across 28 participants, they ran tasks from naturalistic storytelling to simple word repetition, separating the moments of listening from the moments of speaking. Then they did the part that correlation alone cannot do.
In 13 patients they passed small currents of electricity through the electrodes, at carefully chosen timings, and watched what happened to speech. And they went back over the records of hundreds of patients who had had tumours removed from the temporal lobe, to see whose speech suffered when the surgeon’s scalpel passed through which millimetre. Three methods, one question: what is this tissue for?
The answer arrived about a quarter of a second early. In the left posterior superior temporal gyrus — lpSTG, a discrete subzone inside what everyone calls Wernicke’s area — populations of neurons increased their firing roughly 250 milliseconds before a word was spoken, and they did it in continuous speech, ahead of each word rather than in response to the sound of the last one.
The activity was separable from the auditory-feedback signals in neighbouring tissue, which behave the way a proofreader would: they answer the voice. The lpSTG was doing something else. It was already holding the word. Its cells were tuned not to individual sounds or to the movements of the tongue and jaw, but to properties of whole words — how long they are, how many syllables they carry, how often the language uses them.
“It’s more active when you’re about to say the word caterpillar than when you’re about to say the word cat,” said Matthew Leonard, a cognitive neuroscientist at UCSF and co-first author. The lead author, Alexander Silva, put it more sharply: what the region holds is an “auditory image” of the word you are about to say. Wernicke’s phrase, it turns out, was not a metaphor. It was a description of the file format.
Correlation is a description; causation is a verdict, and the verdict came from the operating room. When the researchers stimulated the lpSTG in the moment just before a patient began to speak, the patients made word errors that looked like Wernicke’s aphasia — substitutions and near-misses, the classic signature. Stimulation elsewhere, or at the wrong moment, did not produce the same effect.
The surgical records pointed the same way: patients whose tumour removal had encroached on this subzone were significantly more likely to come out of the operation with speech problems, even though the same operation spares tissue that merely listens. “We already knew that damage to Wernicke’s area can lead to speech production errors, but this publication answers why,” Adeen Flinker, a neuroscientist at NYU who wrote the accompanying commentary in Science and was not involved in the work, told Live Science.
The single-neuron recordings added the strangest detail of all. Some cells in the lpSTG fired selectively for particular words, and the same cells came alive both when the patient said the word and when the patient heard someone else say it. The machinery that builds a word and the machinery that recognises one are not merely neighbours. They overlap.
That overlap is where the clinical stakes are. Speech neuroprostheses — the implants that have let paralysed patients type and speak again in the past few years — have mostly read signals from the motor cortex, the last relay station before the muscles. That works for people whose damage sits below the cortex: ALS, brainstem injury.
It does not work for the far larger population with aphasia, whose language cortex itself is injured; the signal they need has already been scrambled by the time it reaches the motor areas. If speech is planned upstream, in a region that also listens, then decoding there could reach patients that motor-based devices cannot. The team tested the idea in volunteers with severe paralysis after brainstem strokes, who had electrodes implanted for a speech brain-computer interface.
Reading from the lpSTG, the researchers decoded words the participants were trying to say — and, in the same signal, words they were hearing. “What if you could target a single brain region and gain access to the full spectrum of speech, from listening all the way up to speech production?” Silva said. Chang went further at a press briefing, in terms that are as much about human origins as about medicine: this small subzone that turns the words we hear into words we can say may be part of what makes us human.
What the paper does not show is as important as what it does. The paralysis demonstration rests on a handful of participants — the university’s release describes one, the news accounts and the paper report two — and the causal evidence comes from stimulation that lasts seconds inside an operating theatre, not from a chronic implant.
Nobody has yet shown that lpSTG signals stay stable for months or years, which is the entire engineering problem of a neuroprosthesis; the decoding proves a principle, not a product. The lesion evidence is retrospective, drawn from patients who were already having brain surgery for tumours, and the frequency finding is already being described in opposite directions by careful reporters, one writing that activity rose for uncommon words, another for familiar ones — a reminder that press coverage of a 20-page paper is a summary of a summary.
More interesting is the question the paper opens in the frontal lobe. If the main action in both comprehension and production happens in the temporal cortex, what is Broca’s area for? “It is still clearly important,” Leonard said. “In some ways, it’s sort of the next mystery.” Evelina Fedorenko, a neuroscientist at MIT who was not involved, asked the question more bluntly: if both halves of language are happening in the temporal lobe, why do we have the frontal components at all?
Wernicke, it seems, was right about the junction and wrong about the single function — and the brain has one fewer clean room than the textbooks promised, which is the usual price of finally looking at it closely. Every word you said today began, about a quarter of a second before you said it, in a patch of tissue the size of a coin above your left ear. You simply were not in the habit of noticing.
• The paper: Silva, Leonard, Hullett, Levy, Wang, Xu, Bhaya-Grossman & Chang, “A neural code for generative speech production in human Wernicke’s area,” Science 394(6826), p. 858, published 19 November 2026. DOI 10.1126/science.aef8963. Commentary: Adeen Flinker (NYU), “Pinpointing planned speech,” same issue, p. 841. Results were also presented at the Society for Neuroscience annual meeting on 15 November 2026.
• Methods: 28 participants with epilepsy or tumours, already implanted with intracranial electrodes for clinical monitoring; high-density electrocorticography plus single-neuron (Neuropixels) recordings; direct electrical stimulation in 13 patients; retrospective lesion analysis of hundreds of patients who had temporal-lobe tumour surgery.
• The finding: a discrete subzone of the left posterior superior temporal gyrus (lpSTG) becomes active about 250 ms before word onset in continuous speech, ahead of motor cortex activity, and encodes whole-word properties — length, syllables, frequency — rather than phonemes or articulation. It is distinct from adjacent auditory-feedback responses, and many of the same neurons respond to both speaking and hearing the word.
• Causality: stimulation just before speech produced errors resembling Wernicke’s aphasia; resections that damaged the subzone predicted post-operative speech deficits. Both are acute, intraoperative evidence.
• The BCI claim: in volunteers with speech paralysis from brainstem strokes, lpSTG activity decoded both intended and perceived speech. The UCSF release describes one participant; the paper and news accounts report two. Chronic stability and accuracy have not been demonstrated.
• Reported discrepancies: the direction of the frequency effect is described differently by press accounts (Scientific American: more active for uncommon words; National Geographic: stronger for more familiar ones). The paper’s abstract states that the region codes word length and frequency. The subzone’s size is reported as ranging from a bean to a nickel, varying between individuals.
• Still open: the role of Broca’s area; whether the same blueprint applies to reading, writing and typing; long-term implant viability; and whether the subzone explains why some people with damage in this area never develop aphasia.
• Sources: Science (DOI 10.1126/science.aef8963 and 10.1126/science.aem7976); UCSF release (19 Nov 2026); Live Science; Science News; Scientific American; National Geographic.
[심층 리포트] 말보다 0.25초 먼저 도착하는 단어 — 베르니케 영역은 ‘듣는 곳’이 아니라 ‘만드는 곳’이었다
1874년 독일 브레슬라우. 신경해부학자 카를 베르니케는 이상하게 닮은 환자들을 마주 앉혔다. 모두 왼쪽 귀 뒤쪽, 같은 띠 모양의 뇌 조직이 다쳤다. 청력은 멀쩡했다. 말도 했다. 문장은 유창했고 속도도 정상이었으며 억양도 살아 있었다. 문제는 내용이었다. 신발을 말하려던 입에서 ‘발’이 나왔고, 다리를 뜻했던 입에서 ‘못(peg)’이 튀어나왔다.
Korean Version: [심층 리포트] 말보다 0.25초 먼저 도착하는 단어 — 베르니케 영역은 ‘듣는 곳’이 아니라 ‘만드는 곳’이었다
Socko/Ghost