Key Takeaways:
- A UCSF-developed AI system now decodes brain activity into text with 97% accuracy, offering hope for patients who’ve lost speech.
- The technology, tested in epilepsy patients, uses existing brain implants to translate neural patterns into words in real-time.
- Challenges remain, including personalizing models for individual users and reducing error rates before clinical use.
Scientists at the University of California at San Francisco (USCF) have taken another step toward translating brain activity into words, which would allow patients who can no longer speak to communicate again.
Artificial Intelligence For Clinical Trials
Many diseases can deprive patients of speech, despite intact brain capacity. To help them, researchers have been working for years on a “speech decoder” that would translate thoughts into words thanks to a computer connected to the brain implants.
Translating thoughts into words
“Ten years after the first decoding of speech from signals in the human brain, the accuracy, and speed of artificial speech is still much lower than that of natural speech,” said Joseph Makin, David Moses, and Edward Chang of the University of California at San Francisco (USCF). Their work, which is presented in Nature Neuroscience magazine, aims to make the process more fluid. Thanks to the new artificial intelligence (AI) system they have developed, the cortical activity in the brain can now be decoded in real-time and with an accuracy of 97%. “The average word error rate is no more than 3%”, scientists congratulate themselves. “We haven’t reached that point yet, but we think it could be the basis for a speech prosthesis,” says Dr. Joseph Makin.
To achieve these results, his team recruited four Americans and then equipped them with brain implants originally designed to monitor epileptic seizures. The patients then repeated a series of 30 to 50 sentences with up to 250 different words out loud for about 40 minutes. The data collected for each sentence was entered into a machine learning algorithm, which converted them into a series of numbers, which in turn were converted into English phrases.
Eventually, some of them will be used to equip healthy people with this type of brain implant. Elon Musk has invested $150 million to develop his neuralink device for this purpose, while the European Commission is funding the BrainCom project.
FAQ: AI-Powered Speech Decoding
1. What does this AI system do?
It translates brain activity into text, helping people who have lost their ability to speak and communicate again.
2. How does it work?
Brain implants record neural activity while an AI algorithm deciphers it into words, producing real-time text.
3. Who was involved in the study?
Researchers tested it on epilepsy patients with brain implants, having them repeat sentences to train the system.
4. How accurate is it?
It converts speech with 97% accuracy, with an average error rate of 3%.
5. What challenges remain?
Personalizing the system for individuals and ensuring it works reliably outside of lab settings are major hurdles.
6. Is this a cure for speech loss?
No, but it’s a step toward assistive devices that could help non-verbal patients communicate more effectively.
7. How is this different from previous methods?
Earlier systems were slower and less precise. This AI-driven approach improves both speed and accuracy.
8. Is Neuralink involved?
Not directly, but companies like Neuralink and projects like BrainCom aim to develop similar brain-computer interfaces.
9. When will this be available?
It’s still in research stages, so clinical use is likely years away. Further testing and refinement are needed.
Related Reading:
- Researchers at Harvard Show How Our Brains Organize Information About Smells
- How Artificial Intelligence Can Improve Clinical Research
- Artificial Intelligence Can Help Detect Alzheimer’s Six Years Earlier
Bottom Line
This research bridges a critical gap in assistive tech, but perfection is distant. As Dr. Makin cautions, “It’s a promising tool, not a cure-all—yet.”
References
Makin, J.G., Moses, D.A. & Chang, E.F. Machine translation of cortical activity to text with an encoder–decoder framework. Nat Neurosci 23, 575–582 (2020). https://doi.org/10.1038/s41593-020-0608-8
FEEDBACK:
