From: University of Southern California
Novel neural net recognizes spoken words better than human listeners
Machine demonstrates superhuman speech recognition abilities. University of Southern California biomedical engineers have created the world's first machine system that can recognize spoken words better than humans can. A fundamental rethinking of a long-underperforming computer architecture led to their achievement.
The system might soon facilitate voice control of computers and other machines, help the deaf, aid air traffic controllers and others who must understand speech in noisy environments, and instantly produce clean transcripts of conversations, identifying each of the speakers. The U.S. Navy, which listens for the sounds of submarines in the hubbub of the open seas, is another possible user. Potentially, the system's novel underlying principles could have applications in such medical areas as patient monitoring and the reading of electrocardiograms.
In benchmark testing using just a few spoken words, USC's Berger-Liaw Neural Network Speaker Independent Speech Recognition System not only bested all existing computer speech recognition systems but outperformed the keenest human ears.
Neural nets are computing devices that mimic the way brains process information. Speaker-independent systems can recognize a word no matter who or what pronounces it. No previous speaker-independent computer system has ever outperformed humans in recognizing spoken language, even in very small test bases, says system co-designer Theodore W. Berger, Ph.D., a professor of biomedical engineering in the USC School of Engineering.
The system can distinguished words in vast amounts of random "white" noise, noise with amplitude 1,000 times the strength of the target auditory signal. Human listeners can deal with only a fraction as much. And the system can pluck words from the background clutter of other voices, the hubbub heard in bus stations, theater lobbies and cocktail parties, for example. Even the best existing systems fail completely when as little as 10 percent of hubbub masks a speaker's voice. At slightly higher noise levels, the likelihood that a human listener can identify spoken test words is mere chance. By contrast, Berger and Liaw's system functions at 60 percent recognition with a hubbub level 560 times the strength of the target stimulus. With just a minor adjustment, the system can identify different speakers of the same word with superhuman acuity.
Berger and system co-designer Jim-Shih Liaw, Ph.D., achieved this improved performance by paying closer attention to the signal characteristics used by real flesh-and-blood brains in processing information.
First proposed in the 1940s and the subject of intensive research in the '80s and early '90s, neural nets are computers configured to imitate the brain's system of information processing, wherein data are structured not by a central processing unit but by an interlinked network of simple units called neurons. Rather than being programmed, neural nets learn to do tasks through a training regimen in which desired responses to stimuli are reinforced and unwanted ones are not.
"Though mathematical theorists demonstrated that nets should be highly effective for certain kinds of computation (particularly pattern recognition), it has been difficult for artificial neural networks even to approach the power of biological systems," said Liaw, director of the Laboratory for Neural Dynamics and a research assistant professor of biomedical engineering at the USC School of Engineering.
"Even large nets with more than 1,000 neurons and 10,000 interconnections have shown lackluster results compared with theoretical capabilities. Deficiencies were often laid to the fact that even 1,000-neuron networks are tiny, compared with the millions or billions of neurons in biological systems." Remarkably, USC's neural net system uses an architecture consisting of just 11 neurons connected by a mere 30 links.
According to Berger, who has spent years studying biological data-processing systems, previous computer neural nets went wrong by oversimplifying their biological models, omitting a crucial dimension.
"Neurons process information structured in time," he explained. "They communicate with one another in a 'language' whereby the 'meaning' imparted to the receiving neuron is coded into the signal's timing. A pair of pulses separated by a certain time interval excites a certain neuron, while a pair of pulses separated by a shorter or longer interval inhibits it. "So far," Berger continued, "efforts to create neural networks have had silicon neurons transmitting only discreet signals of varying intensity, all clocked the way a computer is clocked, in beats of unvarying duration. But in living cells, the temporal dimension, both in the exciting signal and in the response, is as important as the intensity."
Berger and Liaw created computer chip neurons that closely mimic the signaling behavior of living cells, those of the hippocampus, the brain structure involved in associative learning. "You might say, we let our cells hear the music," Berger said. Berger and Liaw's computer chip neurons were combined into a small neural network using standard architecture. While all the neurons shared the same hippocampus-mimicking general characteristics, each was randomly given slightly different individual characteristics, in much the same way that individual hippocampus neurons would have slightly different individual characteristics. The network created was then trained, using a procedure as unique as the neurons , again taken from the biological model, a learning rule that allows the temporal properties of the net connections to change.
The USC research was funded by the Office of Naval Research; the Defense Department's Advanced Research Projects Agency; the National Centers for Research Resources, and the National Institute of Mental Health. The university has applied for a patent on the system and the architectural concepts on which it is based.
RealVideo demonstration at: