Industries / Frontier science / Brain-computer interfaces
Decoding signals
Decoding is the software that turns activity recorded by implanted electrodes into a cursor movement, a robot arm's motion, text or a voice.
7
Milestones achieved, from 1986 to 2025.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1986
Many cells point the way
Apostolos Georgopoulos, Andrew Schwartz and Ronald Kettner recorded nerve cells in the arm area of a monkey's motor cortex, the strip of brain that directs movement. Each cell fired most for one direction, but only loosely. When each cell's preferred direction was weighted by how strongly it fired and the results were added up, this "population vector" pointed the way the arm moved. Reading many cells together, rather than one, became the basis of movement decoders.
Credit
A. P. Georgopoulos, A. B. Schwartz and R. E. Kettner, in Science
2006
Cursor, then robot arm
In the BrainGate pilot trial, a man with tetraplegia (paralysis of all four limbs) had a 96-electrode array placed in his motor cortex three years after a spinal cord injury. Decoders turned his intended hand movements into a cursor that he used to open simulated e-mail and work a television, even while talking, and he opened and closed a prosthetic hand. In 2012 the same group reported two people with tetraplegia reaching and grasping with a robot arm; one of them, implanted five years earlier, drank coffee from a bottle with it.
Credit
L. R. Hochberg, J. P. Donoghue and colleagues, in Nature
2015
Decoders that recalibrate themselves
Signals from implanted electrodes change over hours and days, so a fixed decoder slowly loses accuracy, and users had to stop and run calibration tasks. Beata Jarosiewicz and BrainGate colleagues tracked the signal's statistics during pauses and retrained the decoder in the background, using the targets the user picked while typing as the answers. People with tetraplegia then typed at their own pace for hours to days without their control getting worse.
Lessonrecorded brain signals do not stay still, so a decoder has to be maintained, not trained once.
Credit
B. Jarosiewicz, L. R. Hochberg and colleagues, in Science Translational Medicine
2021
Handwriting becomes text
A man whose hand was paralysed by a spinal cord injury imagined writing letters by hand. A recurrent neural network, a type of AI model built for sequences, turned his motor cortex activity into text at 90 characters a minute with 94.1% raw accuracy, and over 99% accuracy offline with an autocorrect. The authors give 115 characters a minute as typical smartphone typing for people his age.
Credit
F. R. Willett, J. M. Henderson, K. V. Shenoy and colleagues, in Nature
2024
Speech decoding for daily use
In 2023 two teams published speech decoders in Nature. Stanford's, reading electrode arrays inside the cortex, reached 62 words a minute with a 23.8% word error rate (the share of words wrong) on a 125,000-word vocabulary. UCSF's, reading a grid of electrodes on the brain's surface, reached a median of 78 words a minute with a 25% error rate. In 2024 a UC Davis team reported a man with ALS (amyotrophic lateral sclerosis, a disease that destroys the nerves that work the muscles) whose four arrays gave 97.5% accuracy over 8.4 months; he used it in his own conversations at about 32 words a minute for more than 248 hours.
pmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.govpmc.ncbi.nlm.nih.gov
Credit
F. R. Willett, J. M. Henderson and colleagues (Stanford); S. L. Metzger, E. F. Chang and colleagues (UCSF); N. S. Card, S. D. Stavisky, D. M. Brandman and colleagues (UC Davis)
2025
A voice without delay
The UC Davis team turned the same man's attempted speech straight into sound instead of text. A transformer, the type of AI model behind most chatbots, produced his voice from neural activity within 10 ms, and he could change its pitch and sing short melodies. Listeners matched 94% of sentences to the right transcript out of six choices; when they wrote down what they heard, the median word error rate was 43.75%. UCSF's team separately reported speech synthesis streamed in 80 ms steps from surface electrodes.
Credit
M. Wairagkar, D. M. Brandman, S. D. Stavisky and colleagues, in Nature; K. T. Littlejohn, G. K. Anumanchipalli, E. F. Chang and colleagues, in Nature Neuroscience
2025
Inner speech can be read
A team led from Stanford found that silently imagined speech also shows up in the motor cortex, and decoded imagined sentences in real time in four participants; with a 125,000-word vocabulary, two of them reached word error rates of 26% to 54%. Some unprompted inner speech, such as silent counting, could also be partly decoded. The team tested two safeguards: training decoders to ignore inner speech, and a lock that stays shut until the user silently thinks a keyword ("chitty chitty bang bang"), which was detected correctly 98.75% of the time in real-time tests.
Lessona speech decoder can pick up words the user did not mean to say, so privacy protection has to be built into the decoder.
Credit
E. M. Kunz, F. R. Willett and colleagues, in Cell
Watch for
Decoders that skip recalibration
Neuralink says its trial participants have streamed over 50,000 hours of unlabelled neural data in two years. It says encoders pretrained on each participant's own recordings kept decoders working for weeks without recalibration, and that one participant reached 11.32 bits per second on its cursor task, which it calls a new BCI record, beating its previous 10.39. Bits per second here measures how much a user can select on a grid each second, counting both speed and errors.
Achieved whenthe method and results appear in a peer-reviewed paper, or another group reproduces the gain in decoder lifetime with the test and comparison decoders defined.
Claimed by
Neuralink, October 2026
One pretrained decoder for many
Several groups train single "foundation" decoders on recordings pooled across people and tasks. A preprint by Yizi Zhang, Francis Willett, Liam Paninski and colleagues reports new best scores on the Brain-to-Text '24 and '25 benchmarks with an encoder pretrained across tasks and across species. CorticAll, a NeurIPS 2026 paper trained on about 173 hours from 7 participants, reports the best error rates on 7 of 8 datasets with published baselines. A separate NeurIPS 2026 comparison on up to 345 hours of recordings found that pretraining gains "are often comparable to carefully tuned task-specific training".
Achieved whenan independent comparison shows a pretrained decoder beating carefully tuned one-person decoders on people it was not trained on, in live use.
Claimed by
Zhang and colleagues, November 2025 (revised May 2026); the CorticAll authors, 2026
Open
01
Decoder energy on real chips
An open table that runs the same five or more movement decoders (a Kalman filter, a linear filter, a small recurrent network, a spiking neural network and a small transformer) on at least three low-power chips sold today, and measures accuracy, delay and the energy each decode uses with a power meter. All runs use recordings already published from two monkeys, Indy and Loco, in the NeuroBench "primate reaching" task, so there is no new animal work and no human data. An implant has a small power budget, so these numbers decide which decoders could ever run inside one.
The briefCredit · unclaimed02
Practice data for speech decoders
An open generator that makes fully synthetic recordings of attempted speech from any typed text, in the same form speech decoders read (spike counts and spike-band power on 256 channels), with the faults of real data: noisy channels, drift from day to day and uneven speaking pace. It is checked by training published decoder recipes on it and testing whether they rank in the same order as their published scores on the Brain-to-Text '24 benchmark, so it can be judged without anyone touching patient recordings.
The briefCredit · unclaimed
Open 01 · the brief
Decoder energy on real chips
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open table that runs the same five or more movement decoders (a Kalman filter, a linear filter, a small recurrent network, a spiking neural network and a small transformer) on at least three low-power chips sold today, and measures accuracy, delay and the energy each decode uses with a power meter. All runs use recordings already published from two monkeys, Indy and Loco, in the NeuroBench "primate reaching" task, so there is no new animal work and no human data. An implant has a small power budget, so these numbers decide which decoders could ever run inside one.
Why it is still open
NeuroBench scores this task by counting operations, which its authors say are "measured independently of the underlying hardware" and do not "explicitly correlate with post-deployment latency or energy consumption"; its measured hardware track covers sound classification and an optimisation problem, not decoding. A 2024 TU Delft benchmark of decoders for this data also used counts, not measurements. The one measured result found is a single spiking decoder on one RISC-V microcontroller, GAP9: 1.88 µJ per decode, and 0.50 mW on average when duty-cycled.
Built with what exists
NeuroBench's open code and primate reaching sessions; PyTorch; Google's open-source Coral Compiler v1.0 and the Synaptics Coralboard SL, a board with Google's Coral NPU (neural processing unit) that Google says is on sale worldwide from DigiKey and targets 10 mW average power; the Arduino Nano 33 BLE (Arm Cortex-M4) that NeuroBench itself used as its microcontroller baseline; a bench power monitor. Small, cheap AI chips with open compilers are the advance that makes this a desk job. The site's Note "An open compiler" covers the Coral Compiler.
First step
Run NeuroBench's primate reaching baselines in PyTorch on a laptop, then compile the smallest one with the Coral Compiler and run it in the compiler's simulator.
Done looks like
A public repository and table of at least 5 decoders on at least 3 boards, each row giving R² (NeuroBench's accuracy score, where 1 is perfect) on the six benchmark sessions, latency in milliseconds and measured energy per decode in microjoules, with the measurement setup photographed and scripts that repeat it.
Open 02 · the brief
Practice data for speech decoders
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
An open generator that makes fully synthetic recordings of attempted speech from any typed text, in the same form speech decoders read (spike counts and spike-band power on 256 channels), with the faults of real data: noisy channels, drift from day to day and uneven speaking pace. It is checked by training published decoder recipes on it and testing whether they rank in the same order as their published scores on the Brain-to-Text '24 benchmark, so it can be judged without anyone touching patient recordings.
Why it is still open
BRAND, a real-time BCI platform, includes a speech simulator that maps audio to firing rates with a simple tuning model, built to test speed, not to rank decoders. bcisimulator covers cursor and hand tasks, and its README says its simulators "are not accurate models". A 2026 survey of synthetic brain data benchmarked EEG tasks (motor imagery, seizure detection, steady-state visual responses, auditory attention), not implanted speech recordings. No test of decoder rankings on synthetic speech data was found.
Built with what exists
Python and PyTorch; a pronunciation dictionary to turn text into speech sounds; the published Brain-to-Text '24 entries, whose word error rates run from the 9.7% baseline to the 5.8% winner; AI coding agents to reimplement each published decoder recipe from its paper.
First step
Write a generator for a 50-word vocabulary from the tuning and noise figures given in published papers, and train the Brain-to-Text '24 baseline recipe on it.
Done looks like
A public generator and a synthetic set of at least 10,000 sentences; at least 5 published Brain-to-Text '24 decoder recipes trained and scored on it; their order by word error rate on synthetic data matching their published order for the top three, with the rank correlation and all code released.