💬 A New Algorithm Decodes Thoughts into Speech in Near Real-Time
Scientists from the University of California, San Francisco, tested their technology on Ann, the patient, who has been paralyzed for almost 20 years after a stroke. In 2023, the same team of researchers implanted a device on the surface of the brain's speech region to record the neurons' activity.
An algorithm then converted these signals into real speech. Initial experiments two years ago allowed Ann to speak at a rate of up to 80 words per minute. This is about twice as slow as a natural conversation (about 160 words per minute) but times faster than her current assistant communication device. However, there was almost an eight-second delay between her thoughts and the sounds, which made the speech look more like messaging than a conversation.
The new method reduces this delay between an intent to speak and the words to less than one second. It allows one to communicate in a much more natural format. This was achieved by a speech decoding algorithm similar to those used in voice assistants such as Alexa or Siri, says senior study author Gopala Anumanchipalli.
The AI model that streams Ann's voice was trained on recordings from her wedding video made before the stroke. The model even synthesized words that were not part of the training dataset vocabulary.
🧠 One of the key research team members, neurosurgeon Edward Chang, recently became the head of Echo Neurotechnologies Corp., a company developing implants to connect the brain to a computer. Here, we told more about this startup, which had recently received $50 million in funding.
