Stanford U’s Brain-Computer Interface Enables Stroke and ALS Patients to ‘Speak’ 62 Words per Minute

From drug discovery and protein folding to tumour detection, AI is revolutionizing the biomedical and healthcare fields. Recent research into brain-computer interfaces (BCIs) has revealed their potential to restore rapid communication to people with paralysis by capturing neural activities evoked by attempted speaking actions and decoding these into text. Among the many teams working on such BCIs, Frank Willett’s is one to watch.

Willett is a research scientist at Stanford University’s Howard Hughes Medical Institute. In May 2021, his team was the first to decipher the brain activity associated with handwriting, a breakthrough that enabled people with paralysis to type at a fair pace without using their hands. The team developed an intracortical BCI system that decodes the imaginary handwriting movements of paralyzed patients from neural activity in the motor cortex and translates these handwriting movements into text in real-time using a recurrent neural network (RNN) decoding method. Their research was published as a cover story in the respected journal Nature.

This week, the team advanced their research with the paper A High-performance Speech Neuroprosthesis, introducing a speech BCI that translates speech-related neural activity into text. Theirs is the first speech BCI to record impulse activity from intracortical microelectrode arrays and could benefit people unable to produce clear utterances due to diseases such as stroke and ALS (amyotrophic lateral sclerosis).

The team’s empirical study was conducted with a test subject with ALS who retains the ability to vocalize (unintelligibly) when attempting to speak. Each day, the subject attempted to vocalize 260-480 sentences displayed on a screen while the model recorded spiking activity from four intracortical microelectrode arrays implanted in the left hemisphere of their brain. A recurrent neural network (RNN) was then trained on this data using adapted machine learning speech recognition approaches.

The team used unique input layers for each day to account for across-day changes in the subject’s neural activity and a rolling feature adaptation to account for within-day changes.

The team’s experimental results show that their approach can enable people with speech impairments to communicate at a rate of up to 62 words per minute, which is 3.4 times faster than the previous state-of-the-art speech BCI and approaches the speed of natural conversation (160 words per minute).

The approach also set a new SOTA for accuracy, achieving a word error rate of just 9.1 percent on a vocabulary of 50 words (the previous SOTA speech BCI had an error rate 2.7 times higher) and 23.8 percent on a vocabulary of 125,000 words (the first successful demonstration of large vocabulary decoding).

Overall, this work validates using intracortical speech BCIs to decode attempted speaking movements as a promising approach for restoring rapid communication in people with neurological disorders such as stroke and ALS.

The paper A High-performance Speech Neuroprosthesis is on the bioRxiv repository.

---

Author: Hecate He | Editor: Michael Sarazen

---

We know you don’t want to miss any news or research breakthroughs. Subscribe to our popular newsletter Synced Global AI Weekly to get weekly AI updates.