Casey Harrell: first power user of speech BCI, 3,800+ hours

Casey Harrell has used an implanted brain-computer interface for thousands of hours at home to convert attempted speech into text and to control a computer cursor. Within the first 22.6 months after the implant he logged more than 3,800 hours of independent use at home, the team reported in Nature Medicine, and he has used the system to hold conversations, send messages and continue his work despite paralysis from ALS.

What the system is and how Harrell uses it

Surgeons implanted four microelectrode arrays, each with 64 electrodes, into Harrell’s speech motor cortex during a five-hour operation in July 2023. Each pair of arrays is wired to a pedestal on the exterior of his skull, creating two docking locations to connect the electrodes to a computer. The system records intracortical neural activity from the ventral precentral gyrus and routes that data into parallel decoders for speech and cursor control.

The speech decoder maps neural activity to the 39 phonemes of American English, then converts phonemes to words. The team got a working speech decoder on the first day they began using the device, and on that first session in August Harrell used a 50-word vocabulary with 99.6 percent of words decoded as he intended. The vocabulary was later expanded to 125,000 words; in formal prompted-word tests the research team measured up to about 99 percent word accuracy, and the group reports 97.5 percent accuracy for the 125,000-word vocabulary in the MIT Technology Review account. The Nature Medicine paper gives a prompted-word copy task word accuracy of 99.2 percent for the transformer-based brain-to-text model.

Harrell also uses a separate decoder for cursor control. That enables him to move a cursor and click, so he can send text messages and emails, browse the web and participate in his job as an environmental activist. The system includes software features Harrell requested, such as a privacy mode that deletes decoded text and a profanity filter for conversations with his young daughter.

Independent, long-term use and metrics

Early in the trial researchers had to visit Harrell’s home to connect and disconnect the system. The team automated much of that process so that a care partner can now don and doff the hardware and initialize the software, allowing Harrell to ‘‘wake up, get plugged in, and just get going,’’ according to team member Sergey Stavisky.

The Nature Medicine paper quantifies Harrell’s output over nearly two years: 183,060 sentences totaling 1,960,163 words, an average rate of 56 words per minute, and 92 percent of sentences labeled by Harrell as at least mostly correct. Those usage volumes supported full-time employment and sustained interpersonal communication without researchers present for thousands of hours. The group describes new decoding architectures, including a transformer-based brain-to-text decoder and an RNN-based cursor decoder, that reduce daily recalibration and improve stability.

Harrell calls the device "nothing short of revolutionary." He says it has helped him reconnect with friends and family, read to his seven-year-old daughter and share parenting responsibilities with his wife.

Why it matters

The work demonstrates an intracortical BCI used independently in a home setting for extended periods, addressing two practical barriers: the need for reliable long-term decoder performance and the ability for care partners to operate the system without researcher oversight. The device restored conversational-speed communication and digital access for a person with severe dysarthria and paralysis, showing intracortical BCIs can be more than short-term lab demonstrations.

At the same time, researchers and external experts caution that one user’s outcomes do not guarantee similar durability or acceptance across other people with ALS. Scar tissue, brain degeneration and a general aversion to invasive surgery are unresolved risks highlighted by other long-term implant experiences.

What to watch

The team is working toward a brain-to-voice system that would synthesize natural-sounding cadence and intonation directly from brain activity. The next concrete signals to watch are replication of these uptime and accuracy results in additional participants and longer follow-up to confirm stability across years, and whether trials can reduce the amount of hands-on researcher maintenance even further.