AI Breakthrough:Brain Cells Successfully Perform Speech Recognition

In a major AI breakthrough, clusters of human brain cells, known as brain organoids, were linked to a computer in a landmark attempt to accomplish rudimentary speech recognition. Feng Guo of Indiana University Bloomington is in charge of this project, which aims to develop systems that perform artificial intelligence (AI) operations with a lot less energy than conventional silicon chips.

According to Guo, brain organoids are essentially microscopic formations of nerve cells that merge when stem cells undergo specific growth conditions, similar to miniature brains. It takes two to three months to grow these organoids,  resulting in structures a few millimeters wide containing up to 100 million nerve cells, though far fewer than the approximately 100 billion cells found in human brains. These organoids are then put on a microelectrode array known as “Brainoware,” which allows electrical impulses to be sent to the organoids as well as nerve cell activity detection.

The organoids were trained to recognize the voice of a single individual from a pool of 240 audio recordings, including eight people uttering Japanese vowel sounds, for the speech recognition challenge. The clips were sent to the organoids as signal sequences structured in spatial patterns.

AI Breakthrough: Brain Cells Successfully Perform Speech Recognition

Initially, the organoids achieved an accuracy rate of 30 to 40%. Their accuracy increased dramatically to 70 to 80 percent after two days of training without feedback. Guo refers to this as adaptive learning, emphasizing that the absence of new nerve cell connections caused by a medication did not result in improvement.

Guo’s team’s technique, known as unsupervised learning, involves replaying audio samples without offering input to the organoids on accuracy. This technology tackles two important issues in traditional AI: excessive energy consumption and silicon chip constraints in terms of information and processing separation.

While biocomputing with living nerve cells shows promise, there are still hurdles. Titouan Parcollet of the University of Cambridge acknowledges biocomputing’s possible long-term relevance but cautions against presuming that brain-like components are required to accomplish the current capabilities of deep learning models. Parcollet highlights that current deep-learning models outperform any natural brain in some tasks.

However, Guo’s approach, which focuses on recognizing speakers rather than analyzing speech content, has limits in terms of voice recognition. Furthermore, the lifespan of the brain organoids is limited to one or two months, which is a key barrier that Guo’s team is actively seeking to overcome in order to fully use the computational potential of organoids for AI computing.

ALSO READ: Google Unveils Gemini: AI’s New Era with Advanced Reasoning