An accurate and rapidly calibrating speech neuroprosthesis

Abstract

AbstractBrain-computer interfaces (BCIs) can provide a rapid, intuitive way for people with paralysis to communicate by transforming the cortical activity associated with attempted speech into text. Despite recent advances, communication with BCIs has been restricted by requiring many weeks of training data, and by inadequate decoding accuracy. Here we report a speech BCI that decodes neural activity from 256 microelectrodes in the left precentral gyrus of a person with ALS and severe dysarthria. This system achieves daily word error rates as low as 1% (2.66% average; 9 times fewer errors than previous state-of-the-art speech BCIs) using a comprehensive 125,000-word vocabulary. On the first day of system use, following only 30 minutes of attempted speech training data, the BCI achieved 99.6% word accuracy with a 50 word vocabulary. On the second day of use, we increased the vocabulary size to 125,000 words and after an additional 1.4 hours of training data, the BCI achieved 90.2% word accuracy. At the beginning of subsequent days of use, the BCI reliably achieved 95% word accuracy, and adaptive online fine-tuning continuously improved this accuracy throughout the day. Our participant used the speech BCI in self-paced conversation for over 32 hours to communicate with friends, family, and colleagues (both in-person and over video chat). These results indicate that speech BCIs have reached a level of performance suitable to restore naturalistic communication to people living with severe dysarthria.