AI Accurately Decodes Speech from Brain Activity
An artificial intelligence can interpret words and sentences from brain activity with impressive, albeit constrained, accuracy.
Based on brief recordings of brain activity, the AI can predict what a person has heard, accurately listing the correct answer within its top 10 choices up to 73% of the time, according to a preliminary study by researchers.
“The AI’s performance exceeded expectations for this stage,” remarked Giovanni Di Liberto, a computer scientist at Trinity College Dublin who was not part of the study.
Developed at Meta, the parent company of Facebook, an AI with potential applications for thousands of individuals worldwide who cannot communicate through speech, typing, or gestures was reported on August 25 in arXiv.org. This includes patients in minimally conscious, locked-in, or unresponsive wakefulness syndrome states.
New AI Approach Could Provide a Non-Invasive Alternative to Brain Surgery for Communication Deficits
However, current technologies often require invasive procedures like brain surgery to implant electrodes. This new approach, according to neuroscientist Jean-Rémi King, a Meta AI researcher at École Normale Supérieure in Paris, could offer a non-invasive alternative to help patients with communication deficits.
King and his team trained a language model on 56,000 hours of speech recordings in 53 languages. This model learned to recognize language features at various levels, from letters and syllables to words and sentences.
They then applied an AI using this model to brain activity data from 169 volunteers who listened to stories while undergoing magnetoencephalography or electroencephalography scans.
AI Achieves Up to 73% Accuracy in Decoding Speech from 3 Seconds of Brain Activity Data Using Magnetoencephalography
The AI decoded what participants heard using just three seconds of brain activity data, achieving up to 73% accuracy in predicting the correct answer among more than 1,000 possibilities with magnetoencephalography.
However, challenges remain, such as the bulky and expensive nature of magnetoencephalography machines, limiting practical application in clinical settings.
Additionally, the current study focused on decoding speech perception rather than production, emphasizing the distance still needed to develop meaningful communication tools for nonverbal patients.
Read the oginal article on: Science News
Read more: Scientists Discover Brain Vascular Changes Associated with Alzheimer’s Disease
