Ann is 47 and had not spoken for 18 years after a brainstem stroke left her paralysed. She can now hold a conversation — not by typing, not by selecting letters with her eyes, but by trying to speak and having the words come out of a computer within a fraction of a second, in a voice reconstructed from recordings made before her injury. The system behind it, from researchers at the University of California, Berkeley and UC San Francisco, is described this week in Nature Neuroscience.
What changed: the delay
Speech implants are not new. The problem with earlier ones was latency: the user would attempt a whole sentence, the system would decode it, and up to eight seconds later the sentence would be spoken. Eight seconds is long enough to kill any conversation. The Berkeley team’s advance is streaming — the same trick voice assistants like Siri and Alexa use to start answering before you have finished asking. Their AI decodes the brain’s signals in short chunks and speaks each as it arrives, so the words flow at close to the pace of natural talk.
How it works
An array of electrodes sits on the surface of the motor cortex, the strip of brain that would normally control the lips, tongue, jaw and larynx. When Ann tries to say a word, those neurons fire even though the muscles no longer respond. The electrodes pick up the pattern; a neural network trained on many hours of her attempts maps it to sounds; a synthesiser speaks them. “The neuroprosthesis takes neural signals from the motor cortex, which controls speech production, and uses AI to translate them into speech,” says lead researcher Cheol Jun Cho. Because the training data included recordings of Ann from before her stroke, the synthesiser produces something close to her own voice rather than a generic one.
Who this is for
The immediate group is people with severe paralysis from brainstem stroke, ALS or spinal injury who retain the intent to speak. The researchers expect the approach to extend to other neurological conditions that take away speech. Bangladesh has a very large stroke population and almost no assistive-speech technology; an implant is years and a great deal of money away from a ward in Dhaka. But the software half of this work — decoding intent quickly enough to hold a conversation — is the part that will travel, and it is the part that just got solved.
Source: Indian Express, via Prothom Alo.
