Speech neuroprosthesis · Grey Matter

A speech neuroprosthesis is an implanted brain-computer interface that reads the cortex controlling the lips, tongue, jaw and larynx while a person tries to speak, and turns that activity into text or a synthetic voice for people who have lost speech to paralysis.

Speech neuroprosthesis. A speech neuroprosthesis is an implanted brain-computer interface that reads the cortex controlling the lips, tongue, jaw and larynx while a person tries to speak, and turns that activity into text or a synthetic voice for people who have lost speech to paralysis.

The electrodes sit on or inside the ventral part of the motor cortex, because that strip keeps firing with the attempt to speak even when the nerves or the muscles can no longer carry it out. A neural decoder turns each short window of activity into probabilities for the roughly forty phonemes of English, and a language model assembles the likeliest sentence. In 2023 one system with four Utah arrays in a woman with ALS reached 62 words per minute with a 23.8 % word error rate on a 125,000-word vocabulary, and another with a 253-electrode electrocorticography grid in a woman paralysed by a brainstem stroke reached 78 words per minute and drove a talking avatar. In 2024 a man with ALS and 256 electrodes reached 97.5 % word accuracy over 8.4 months; by 2026 he had used it on his own at home for more than 3,800 hours, at 56 words per minute.

It reads the attempt to move. The user has to try to say the words, so it suits people whose path to the muscles is broken.

The language model carries much of the accuracy, correcting the errors of the phoneme stage, which helps ordinary sentences more than unusual names.

Calibration has become short. The 2024 system worked on its first day, after 30 minutes of recordings, at 99.6 % on a 50-word vocabulary.

It can catch words that were never attempted, since inner speech leaves a weaker copy of the same pattern in motor cortex; see mental privacy.

A speech neuroprosthesis restores the last step of speaking and depends on every step before it.

The words still have to be chosen and the command to articulate them still has to form in the cortex; the device only replaces the nerves and muscles that would have carried it out.

Questions: How long can a speech implant keep working once it is in? Years, as far as anyone has measured. A man with ALS used four Utah arrays for speech nearly every day for almost two years at home, more than 3,800 hours, and Utah arrays in other people have recorded for over eight years, although each needle slowly loses neurons as scar tissue forms around it. In ALS the limit can also be the disease: a surface implant used for almost seven years faded after the sixth, as the cortex under it shrank, while the hardware kept working. Why do so many speech-decoding studies use recordings from people with epilepsy? They already have electrodes inside the skull for clinical reasons and spend days on a ward waiting for seizures, and many agree to do research tasks in between. Their electrodes are placed to find seizures, so each person covers a different, uneven set of regions, which used to confine every decoder to one patient. A 2025 study in Nature Communications turned that into a strength: it pooled stereo-EEG recordings from 25 people saying tongue twisters aloud, and the decoder trained on the group predicted phonemes better than decoders trained on each person alone. These patients can still speak, so the work maps the speech network and tests methods, while a speech neuroprosthesis for paralysis still needs its own implant. Why are speech implants placed in the motor cortex rather than in Broca's area? Because the motor cortex is where the signal turned out to be. The 2023 speech implant at Stanford put two arrays in the ventral motor cortex and two in area 44, part of Broca's area, and the arrays in area 44 carried almost no information about the movements, sounds or words the participant attempted (below 12 % accuracy in classifying them), so every result came from the motor arrays. Broca's area seems to work at a higher level of planning, spread over a wider network, while the strip of motor cortex that drives the lips, tongue and larynx gives a dense, local pattern for each sound that a few hundred electrodes can read.