Two women who lost the ability to speak clearly used brain implants to communicate through computers at speeds approaching ordinary conversation. The systems interpreted brain activity linked to attempted speech, producing 62 and 78 words per minute in separate studies.
Reading the signals behind attempted speech
A brain-computer interface, or BCI, collects brain signals and turns them into commands for an external device. Earlier systems had allowed people with paralysis to control robotic arms, play video games and send emails. Translating intended speech into text had also been demonstrated, but speed, accuracy and vocabulary were limited.
The new studies used artificial intelligence to connect brain activity with the movements involved in speech. Rather than requiring the women to produce understandable sounds, the systems learned patterns associated with moving the lips, jaw and tongue, then used those patterns to construct words and sentences.
Both women volunteered for brain implants. One, Pat Bennett, has ALS, a disease that affects motor neurons. The other, identified as Ann, had experienced a stroke in her brain stem. They remembered how to form words, although paralysis prevented them from enunciating clearly.
Two approaches, and two different results
Researchers at Stanford used the Utah array, a small square sensor with 64 needle-like bristles tipped with electrodes. A surgeon placed four sensors in Bennett’s cerebral cortex in March 2022. Over four months, researchers asked her to attempt sentences aloud while software learned to decode the neural signals.
The system produced 62 words per minute using a 125,000-word vocabulary. It made errors 23.8 percent of the time. That was substantially faster than the previous record of 18 words per minute, set in 2021 with a system that translated imagined handwriting into text.
A team at UC San Francisco took a different route. Its paper-thin array, with 253 electrodes, rested on the brain’s surface and recorded activity across the speech cortex. Ann moved her lips without making sounds as the researchers trained a deep-learning model on phrases from a 1,024-word conversational vocabulary.
That system decoded speech at 78 words per minute, compared with the 14 words per minute Ann was used to with her type-to-talk device. Its error rate was 4.9 percent on sentences drawn from a 50-phrase set. Simulations estimated a 28 percent word error rate with a vocabulary of more than 39,000 words.
From text on a screen to a computer voice
The UCSF system also went beyond displaying words. Researchers created a digital avatar to speak Ann’s intended words aloud. They gave the animated woman brown hair like Ann’s and used wedding footage to make the avatar’s voice sound like hers.
This design reflects a broader aim: communication can involve personal expression as well as conveying words. The researchers hope that a personalized voice and visual presence could make interactions with family and friends feel more natural and expressive.
The pace is promising, but it remains below the roughly 160-word-per-minute rate of natural conversation among English speakers. The studies show that attempted speech can be decoded much faster than before; they do not establish that the systems are ready for routine, everyday use.
Accuracy and durability remain open challenges
The two electrode designs involve trade-offs. Sensors implanted in the brain can record individual neurons and provide more detailed signals. But they can shift, and even a movement of a millimeter or two can change what they record. Scar tissue may also affect signal quality over time.
A surface array captures less detailed activity, but covers a larger area. Its signals draw on thousands of neurons and may be more stable. Neither approach removes the practical challenge of maintaining useful recordings over weeks, months or years.
The number of electrodes that can safely be placed in the brain also limits how much activity the systems can measure. More electrodes could provide a clearer picture of brain activity, but the current results already mark progress toward restoring faster communication for people with paralysis. Researchers see the work as a step toward speech that is more fluid, reliable and personal.