From drug discovery and protein folding to tumor detection, AI is revolutionizing biomedical and healthcare fields. Recent research on brain-computer interfaces (BCIs) has shown that capturing neural activity triggered by speech attempts and decoding them into text may restore rapid communication with people with paralysis. One thing became clear. Among the many teams working on BCI like this, Frank Willett’s team is his one to watch.
Willett is a researcher at the Howard Hughes Medical Institute at Stanford University. In May 2021, his team deciphered the brain activity associated with handwriting for the first time. This is a breakthrough that allows people with paralysis to type at a reasonable pace without using their hands. The team developed an intracortical BCI system that decodes paralyzed patients’ imaginary handwritten movements from neural activity in the motor cortex and uses recurrent neural network (RNN) decoding to convert these handwritten movements into text in real time. developed.Their research was published as a cover story in a prominent journal Nature.
This week the team advanced their research with a paper High performance voice neuroprosthesis, introduces a speech BCI that converts speech-related neural activity into text. Theirs was the first speech BCI to record impulse activity from an intracortical microelectrode array, allowing for clear speech for diseases such as stroke and he ALS (amyotrophic lateral sclerosis). can be useful for people who cannot
The team’s empirical study was conducted on subjects with ALS who retained the ability to vocalize (unintelligibly) when trying to speak. Each day, the subject attempted to utter 260–480 sentences displayed on the screen, and the model recorded spiking activity from her four intracortical microelectrode arrays implanted in the left hemisphere of the brain. We then trained a recurrent neural network (RNN) on this data using an adapted machine learning speech recognition approach.
The team used a daily-specific input layer to describe changes in subjects’ neural activity over the course of a day, and rolling functional adaptation to describe changes within a day.
The team’s experimental results show that their approach enables people with speech disorders to communicate at rates of up to 62 words per minute. This is 3.4 times faster than the previous state-of-the-art speech BCI, approaching natural speech speed (160 words per minute).
This approach also sets the new SOTA for accuracy, achieving a word error rate of just 9.1% on a 50-word vocabulary (the error rate of the previous SOTA speech BCI was 2.7 times higher), and a 125,000-word We achieved 23.8% in vocabulary (first successful demonstration of large vocabulary decoding).
Overall, this work validates decoding attempted speech behavior using intracortical speech BCI as a promising approach for restoring rapid communication in people with neurological disorders such as stroke and ALS. To do.
paper High performance voice neuroprosthesis It’s in the bioRxiv repository.
author: Hecate He | Editor: Michael Sarazen
We know you don’t want to miss news or research breakthroughs. Subscribe to popular newsletters Synchronized Global AI Weekly Get weekly AI updates.
