A brain-computer interface has gotten better than ever before at translating thoughts from people with speech difficulties. Researchers are also thinking through how to protect users’ privacy
In recent years, scientists have been working on technology that could help people—such as those paralyzed by strokes or with neurological conditions like ALS—to communicate. In particular, devices called brain-computer interfaces, or BCIs, pick up electrical signals in the brain as people try to form words, then translate these signals out loud. But in a new study, published Thursday in the journal Cell, researchers report they have decoded four participants’ inner voices for the first time, with up to 74 percent accuracy.
“It’s a fantastic advance,” Christian Herff, a neuroscientist at Maastricht University in the Netherlands who was not involved in the new study, tells the New York Times’ Carl Zimmer, adding that the technology not only helps people speak who otherwise can’t but also improves scientists’ understanding of how language works.
In previous uses of BCIs, researchers asked patients who couldn’t speak to try to physically form words. This created signals of so-called attempted speech in the brain, which were picked up by implanted electrodes and decoded with a computer algorithm. This process, however, could be tiring and uncomfortable for users, as Erin Kunz, an electrical engineer at Stanford University, tells Science’s Annika Inampudi.
So, Kunz and her team decided to try to investigate whether implanted electrodes could pick up a patient’s inner voice. If it could, operating the device might become easier for users. But this research was also meant to investigate questions around privacy—could these devices also pick up unintentional speech? After all, our inner voices might say things we don’t want others to hear.
The team says that so far, the technology isn’t accurate enough to translate thoughts against a participant’s will. “We’re focused on helping people … who have issues with speaking,” Kunz tells Science.
To protect users’ privacy, they chose a passphrase to activate the device that was unlikely to come up in everyday speech: “Chitty Chitty Bang Bang,” the title of the 1964 Ian Fleming novel and 1968 movie. The technology would start translating thoughts when it detected the phrase, which, for one participant, it did with 98.75 percent accuracy.
In the tests, the researchers asked the participants—all four of whom have some trouble speaking—to either attempt saying a set of seven words or to merely think them. They found the patterns of neural activity and regions of the brain used in both scenarios were similar, but the inner thoughts produced weaker signals.
Then, the team trained the computer system on the signals produced when participants thought words from a 125,000-word vocabulary. When the users then thought sentences with these words, the device translated the resulting brain activity. The technology produced words with an error rate of 26 to 54 percent, making it the most accurate attempt to decode inner speech to date, Science reports.
In another experiment, participants were asked to tally circles on a screen as the device tried to translate their thoughts. This was meant to test whether the computer would pick up internal thoughts that the participants weren’t told to say, and in some cases, the system picked up a number.
Beyond helping people speak who otherwise couldn’t, the findings show that, at least for some people, language plays a role in the process of thought, Herff tells the New York Times. It also reveals neural differences between attempted and internal speech, Silvia Marchesotti, a neuroengineer at the University of Geneva who wasn’t involved in the research, tells Nature’s Gemma Conroy.
But the researchers say they’re not done yet—and they’re working toward even better outcomes.
“The results are an initial proof of concept more than anything,” Kunz tells the New York Times. “We haven’t hit the ceiling yet.”