The world is just getting used to the power and sophistication of virtual assistants created by companies like Yandex and Google, which can decipher our spoken language with eerie accuracy compared to what technology was capable of just a few years ago.
In reality, however, a much more impressive and overwhelming milestone may be just around the corner, making speech recognition almost like a child's game: artificial intelligence (AI) systems that can translate our brain activity into fully formed text without hearing a single word.
This is not really science fiction. In recent decades, brain-machine interfaces have evolved at a rapid pace, from working with animals to human participants, and, in fact, there are already attempts of this kind.
Researchers at the University of California, San Francisco explain this in a new study.
A team of scientists led by neurosurgeon Edward Chang of the UCSF laboratory has used a new method to decode an electrocorticogram: recording electrical impulses that occur during cortical activity recorded by electrodes implanted in the brain.
In a study in which four epileptic patients wore implants to monitor seizures caused by their medical condition, the UCSF team conducted a side-by-side experiment in which participants read and repeated several suggested sentences aloud, while electrodes recorded their brain activity during exercise.
This data was then transmitted to a neural network that analyzed patterns of brain activity corresponding to specific speech signatures, such as vowels, consonants, or mouth movements, based on audio recordings of the experiment.
Then another neural network decoded these representations – gleaned from repetitions of 30-50 spoken sentences – and used it to try to predict what was said based solely on cortical word signatures.
The system showed a word error rate of only 3 percent.
While there are many obstacles to overcome, scientists speculate that the system may one day act as the backbone of a speech prosthesis for patients who have lost the ability to speak.
The results are presented in the journal Nature Neuroscience.
