The patient, Pat Bennett, had lost her ability to communicate through speech due to the neurological disease; however, her brain was still sending signals of her intent to talk. The researchers were able to read these signals and train an algorithm — what they called a Recurrent Neural Network (RNN) decoder — to recognize her attempts to speak. Through a brain-computer interface (BCI), the researchers demonstrated speech-to-text that records spiking activity from intracortical microelectrode arrays at a rate of 62 words per minute, with better than 75% accuracy. The study authors wrote that she, “achieved a 9.1% word error rate on a 50-word vocabulary (2.7 times fewer errors than the previous state-of-the-art speech BCI – Brain Computer Interfaces) and a 23.8% word error rate on a 125,000-word vocabulary (the first successful demonstration, to our knowledge, of large-vocabulary decoding). The participant’s attempted speech was decoded at 62 words per minute, which is 3.4 times as fast as the previous record and begins to approach the speed of natural conversation.” The speed of natural conversation is about 160 words per minute.
During the study, the researchers used the Recurrent Neural Network (RNN) to record data while Bennett attempted to speak 260-480 sentences at her own pace per day that were chosen randomly. The algorithm was then trained on the data from all combined days, “using custom machine learning methods adapted from modern speech recognition to achieve high performance on limited amount of neural data.”
“In particular, we used unique input layers for each day to account for across-day changes in neural activity, and rolling feature adaptation to account for within-day changes,” according to the study. “By the final day our training dataset consisted of 10,850 total sentences. Data collection and RNN training lasted for 140 min per day on average (including breaks).”
While the use of such technology is a feat of its own and the brain activity-assisted algorithm is accurate enough to generally get an understanding of a sentence, the researchers conceded that the error rate is still too high for everyday use and the system is not yet viable for patients. “Our demonstration is a proof of concept that decoding attempted speaking movements with a large vocabulary is possible using neural spiking activity,” the study authors wrote. “However, it is important to note that it does not yet constitute a complete, clinically viable system. Importantly, a 24% word error rate is probably not yet sufficiently low for everyday use. Nevertheless, we believe that our results are promising.”
BCI technology, however, is rapidly being improved upon, fueling researcher’s hope that increasingly positive outcomes could still be possible. Plus, the authors found that the error rate did decrease as they added more electrodes to read the brain’s signals. “Taken together, these findings suggest that a higher channel count system that records from only a small area of 6v [ventral premotor cortex] is a feasible path towards the development of a device that can restore communication at conversational speeds to people with paralysis,” according to the study.
The study authors are not the only ones investing in BCI technology. Currently, companies like Musk’s Neuralink, Paradromics, Precision Neuroscience, and Blackrock Neurotech are all working on devices that incorporate additional electrodes or channels to get higher resolution access to the brain’s signals. Blackrock first implanted its Neuroport Array chip into a human brain all the way back in 2004. The Neuroport device pokes through the skin, and places 96 array needles in the brain to open up 600+ channels of communication and has been implanted more than 50 times. The company is currently working on a higher bandwidth brain link with a 10,000-plus channel interface called the Neuralace. Blackrock plans to make the device available to researchers by 2024. Additionally, earlier this year Neuralink announced it had been approved by FDA to start the first-in-human clinical trial for its N1 BCI.
While experts say the use of this kind of restorative technology is years away from commercial availability, the authors research adds an additional layer to what 50 years ago would have been seen as a Syfy pipedream: restoring the ability to communicate at conversational speeds in people with paralysis.
REFERENCE: MD&DI (Medical Device and Diagnostic Industry); 07 SEP 2023; Katie Hobbins