A man with ALS regained speech at 97 percent accuracy

Researchers in the BrainGate consortium implanted four microelectrode arrays, 256 channels in total, into the part of his brain that controls speech, then trained a machine-learning system to translate the resulting signals directly into words from a 125,000-word vocabulary.

The system reached 97.5 percent word accuracy, and useful performance began on the very first day of use, 25 days after the implant surgery, well before most brain-computer interfaces of this kind typically become usable. He used the system to hold conversations at a rate of about 32 words per minute.

The vocabulary size and accuracy were both larger than prior speech neuroprostheses had achieved for someone with his level of impairment, and the fast calibration meant he did not need weeks of retraining before the device became useful in daily life. He still has some limited natural speech; whether this approach extends to people who have lost speech completely is an open question the same research group is now testing.

Twenty-five days after surgery, he was talking again.