The company shared research on the AI model, which can decod speech from non -invasive records of brain activity. He has the potential to help people after traumatic brain damage, which caused that they could not communicate through speech, writing or gestures.
Decoding of speech based on brain activity was a long -term goal of neuronauts and clinicists, but most of the progress was based on invasive brain recording techniques, such as stereotactic electroencephalography and electrocorticography.
Meanwhile, scientists from Meta believe that the transformation of speech using non -invasive methods will provide a safer, more scalable solution that could ultimately benefit more people. In this way, they created a model of deep learning trained with contrasting learning, and then used it to adapt non -invasive brain recordings and speech sounds.
To do this, scientists used Open Source, a self -sufficient learning model Wave2Vec 2.0 To identify complex speech representations in volunteer brains while listening to audiobooks.
The process includes the contribution of electroencephalphia and magnetoencefalographic recordings to the “brain” model, which consists of a standard deep network of dispute with residual connections. Then the created architecture learns to equalize the results of this brain model to deep representation of speech sounds that were presented to participants.
After the training, the system performs the so -called zero classification: with a fragment of brain activity, it can determine on the basis of a large pool of new audio files that it actually heard.
According to the finish: “The results of our research are encouraging because they show that trained AI can successfully decode speech from non -invasive records of brain activity, despite noise and inherent variability in this data. These results are only the first step, but in this work we focus on the fact that this work may help. Patients potentially include new ways of interaction with computers.”
Learn more about research Here
