Summary: For the first time, scientists have decoded inner speech—the silent words people think—on demand using brain-computer interface (BCI) technology, reaching up to 74% accuracy. By recording neural signals from people with severe paralysis, researchers showed that imagined speech and attempted spoken speech share overlapping patterns in the motor cortex, although inner speech signals are generally weaker. This advance points toward faster, more natural communication options for people who cannot speak aloud, with privacy safeguards and room for improved accuracy as devices and algorithms evolve.
Researchers trained artificial intelligence models on motor-cortex activity linked to inner speech and demonstrated that imagined sentences could be identified from a large vocabulary. The team also implemented a thought-based password system that prevents decoding unless the user intentionally unlocks the BCI, addressing privacy concerns while enabling convenient use.
Key Facts
- Breakthrough decoding: Inner speech was decoded on command with up to 74% accuracy from motor-cortex activity.
- Shared neural patterns: Attempted and imagined speech recruit overlapping regions of the motor cortex, though inner speech tends to have weaker activation.
- Privacy control: A thought-based password can lock and unlock inner-speech decoding, minimizing unwanted interpretation of private thoughts.
Source: Cell Press
Researchers have isolated and decoded neural activity linked to inner speech—the silent internal monologue—and demonstrated on-demand decoding with substantial accuracy.
Published August 14 in the journal Cell, the study outlines how BCIs can begin translating inner thoughts when a participant intentionally triggers decoding with a mental password. The approach could broaden communication options for people with severe speech and motor impairments by making interactions via BCI feel quicker and more natural than systems that require physical attempts to speak.
“This is the first time we’ve managed to understand what brain activity looks like when you just think about speaking,” said lead author Erin Kunz of Stanford University. “For people with severe speech and motor impairments, BCIs capable of decoding inner speech could help them communicate much more easily and more naturally.”
BCIs already decode movement-related activity to control prosthetics and can interpret attempted speech when users engage speech-related muscles without producing audible words. However, attempting speech can be tiring for people with limited muscle function. The team therefore explored whether BCIs could decode inner speech—thinking the words without any physical articulation—to reduce effort and speed up communication.
The researchers recorded neural activity from microelectrodes implanted in the motor cortex of four participants who had severe paralysis due to either amyotrophic lateral sclerosis (ALS) or brainstem stroke. Participants were instructed to either attempt to speak or to imagine saying specific words and sentences. The recordings showed that inner speech and attempted speech activate overlapping motor-cortex regions and produce similar neural patterns, with inner speech typically producing lower-amplitude responses.
Using the inner-speech recordings, the team trained AI models to interpret imagined words and sentences. In a proof-of-concept demonstration, the BCI decoded imagined sentences from vocabularies as large as 125,000 words, with peak accuracy reported at 74%. The system also detected unprompted inner responses—for example, participants silently counting numbers when asked to tally items on a screen.
Crucially, the study identified a neural dimension that distinguishes attempted speech from imagined speech, enabling BCIs to selectively ignore inner speech if desired. To safeguard privacy, the team demonstrated a password-controlled mechanism: users could think a preset phrase to unlock inner-speech decoding. In experiments, the mental password was recognized with better than 98% accuracy, preventing unintended decoding of private thoughts unless the user deliberately activated the system.
The researchers emphasize that current BCIs are not yet capable of reliably decoding unrestricted free-form inner speech without substantial errors. However, they point to clear potential: more sensors, expanded datasets, and improved algorithms could increase fidelity and enable richer, conversational communication for people who cannot speak.
“The future of BCIs is bright,” said senior author Frank Willett of Stanford University. “This work gives real hope that speech BCIs can one day restore communication that is as fluent, natural, and comfortable as conversational speech.”
Funding:
This research was supported by the Assistant Secretary of Defense for Health Affairs, the National Institutes of Health, the Simons Collaboration for the Global Brain, the A.P. Giannini Foundation, Department of Veterans Affairs, the Wu Tsai Neurosciences Institute, the Howard Hughes Medical Institute, Larry and Pamela Garlick, the National Institute on Deafness and Other Communication Disorders, the National Institute of Neurological Disorders and Stroke, the Eunice Kennedy Shriver National Institute of Child Health and Human Development, the Blavatnik Family Foundation, and the National Science Foundation.
About this neurotech and AI research news
Author: Julia Grimmett
Source: Cell Press
Contact: Julia Grimmett – Cell Press
Image: Image credited to Neuroscience News
Original Research: Open access.
Title: “Inner speech in motor cortex and implications for speech neuroprostheses” by Erin Kunz et al., published in Cell.
Abstract
Inner speech in motor cortex and implications for speech neuroprostheses
Speech BCIs hold promise for restoring communication to people with paralysis, but they raise questions about the potential to decode private inner speech. Inner speech also offers a possible alternative to requiring users to physically attempt speech—a process that can be exhausting and slow. Using multi-unit recordings from four participants, the researchers show that inner speech is reliably represented in the motor cortex and that imagined sentences can be decoded in real time. Inner speech representations are highly correlated with attempted speech, yet a distinct neural “motor-intent” component separates the two. The team explored decoding aspects of free-form inner speech during sequence recall and counting tasks and developed high-fidelity strategies to prevent unintentional decoding of private inner speech, demonstrating both capability and control for future speech neuroprostheses.