Researchers from Stanford University have published a paper claiming to record thought-to-text communication via a brain-computer interface (BCI). A rival approaches.
In a preprint that has not yet been peer-reviewed, MIT Technology Review, the team explains the inner workings of its new ‘neuroprosthesis’. This is a brain-computer interface (BCI) that uses an intracortical microelectrode array to capture high-resolution recordings of a user’s brain activity associated with speech.
To prove this concept, the team recruited one research participant. It is an unidentified member of the general public suffering from amyotrophic lateral sclerosis (ALS) (also known as Lou Gehrig’s disease), an illness that has rendered it incapable of producing comprehensible speech. When wearing the neuroprosthesis BCI, the subject thought words and he was able to decode them at a rate of 62 words per minute. This makes him more than three times faster than his previous state-of-the-art BCI voice system.
Of course, speed doesn’t come without accuracy, but here too the team claims a major milestone: with a limited 50-word vocabulary, the system exhibits an error rate of 9.1%, almost 3% faster than its rivals. It was a 1 in 1 error rate. With a vocabulary of 125,000 words, the error rate rose to 23.8% and still proved usable.
However, sensor systems actually focus on movement rather than detecting thoughts associated with speech. It builds on previous work using the same system to control a robotic arm or an on-screen keyboard. Subjects simply attempt to speak, and implanted sensors record brain activity associated with speech-related mouth and face movements, which are then decoded by a specially trained recurrent neural network (RNN) — Even if the user’s mouth doesn’t actually move.
The system works with implanted sensors designed to detect brain activity associated with planned muscle movements. (📷: Willett et al.)
“Our demonstration is a proof of concept that deciphering attempted speech actions from intracortical recordings is a promising approach,” the researchers admit. It trains the decoder and adapts to changes in neural activity that occur over days without the user needing to pause and readjust her BCI. Perhaps most importantly, the 24% word error rate is likely still not low enough for everyday use. ”
preprint available On the bioRxiv server at Cold Spring Harbor Laboratory Now under open access conditions.