Forward-looking: Stanford University researchers have unveiled a brain-computer interface capable of translating imagined words directly from neural activity into speech – marking a first in neurotechnology. Unlike earlier systems that relied on detecting brain signals generated when people tried to move their mouths or vocal cords, the new approach works even when a person simply thinks of speaking.
Four people living with severe paralysis, caused by conditions including amyotrophic lateral sclerosis and brainstem stroke, took part in the trial. One participant could only respond by moving his eyes – up for yes and side-to-side for no. In the study, described this week in Cell, doctors implanted microscopic electrode arrays into each participant's motor cortex, the brain region that normally directs movements involved in speech.
The technology was developed under the BrainGate BCI consortium, a long-running academic collaboration in brain-computer interface research. Once in place, the electrodes recorded activity in the speech motor cortex while participants were presented with two types of tasks: attempting to speak aloud and silently imagining specific words.
Machine learning models were trained to detect and classify distinct patterns of brain activity linked to phonemes – the smallest individual sound units in spoken language. The system then recombined these phonemes into whole words and sentences in real time. The researchers found that imagined speech produced a weaker but still distinct neural signature compared to attempted speech. Even so, the decoding system reached accuracy rates of up to 74 percent.
"This is the first time we've managed to understand what brain activity looks like when you just think about speaking," study lead author Erin Kunz, a neuroscientist at Stanford, told The Financial Times. For individuals with profound speech and motor impairments, she said, BCIs that understand inner speech could make communicating "easier and more natural."
Stanford neurosurgery assistant professor Frank Willett, a senior member of the team, said that the results show how far the field has progressed toward restoring conversational communication to people who cannot speak. Attempting speech, he noted, can be physically draining for those with partial paralysis and may produce unwanted vocalizations or breathing difficulties. Decoding silent speech directly from the brain could eliminate these drawbacks.
The researchers also discovered an important privacy concern. In some cases, the system detected words that participants had not been asked to think about – such as counting numbers during a visual task. To address this, the team created a form of mental lock in which the decoder remains inactive unless triggered by an imagined password. In testing, the phrase "chitty chitty bang bang" successfully blocked unintended decoding 98 percent of the time.
The breakthrough comes amid growing interest in BCIs from both academic and commercial sectors. Investment in the field is expected to intensify following the launch of Merge, a new company backed by OpenAI chief executive Sam Altman, intended to compete with Elon Musk's Neuralink.
While the Stanford work remains experimental, the researchers argue it provides proof-of-principle that future devices could let users speak fluently using thought alone. "This work gives real hope," Willett said, "that speech BCIs can one day restore communication that is as fluent, natural and comfortable as conversational speech."
Stanford's brain-computer interface turns inner speech into spoken words
