Georgie Tech
Machine Hearing
If you've ever worked with speech recognition programs, you've learned that they don't work well in a noisy environment. Say the same thing to a person in the same environment, and he or she understands you just fine. Distorted speech, dropped words— all of that is relatively simple for a human to decipher. Not so for a computer.
Machine Hearing
MAJOR INNOVATION Mixing low-power analog input with analysis of sound.
WHY IT MATTERS Security and voice recognition will be the first apps to benefit.
ESTIMATED ARRIVAL Star Trek–esque voice recognition is five to eight years out.
"I realized, if we're going to really try to mimic human performance in certain tasks, we need more computing power," says David Anderson, associate professor for Digital Signal Processing and Computer Engineering at the Georgia Institute of Technology. Hearing doesn't require as much power as listening, however.
Working with colleague Paul Hasler, Anderson has been experimenting with ultra-low-power processors with analog audio inputs to handle "machine hearing." Using just 5 to 30 microwatts of power, the team has managed to perform signal analysis on a variety of sounds, including human speech. While the technology isn't as flexible as the human ear, it's been good enough to differentiate background noise from what should be heard. Because the input uses so little power, Anderson claims such a sensor could run on an AA battery for a year or two. A small solar cell might keep it going indefinitely.
Using field-programmable analog arrays (FPAAs), he and his team have created prototypes of advanced digital signal processors that work with analog input. "[Digital is] predictable and easy to program," says Anderson. "But you pay a price of power."
Surveillance listening equipment that's little more than a microphone and a machine hearing chip can listen for specific events: breaking glass, gun shots, or passers-by murmuring in the corridors. All those noises can be differentiated. Once a specific sound is detected, the chip alerts a security program.
One of the integrated circuits is a noise-suppression chip that can extract speech signals out of background noise using front-end processing. "It's been very effective to reduce noise without distorting speech," says Anderson. The actual recognition of what's said, however, is another matter.
Machine hearing isn't the same as speech recognition, but the two go together like puzzle pieces. Today, doing full signal analysis of speech a computer hears in real time is difficult. Most analysis is done using all the resources of a single PC, yet it still sometimes takes hours to get results. Anderson says there's a better way: "We always try to emulate the brain." The more the team emulates the gray matter with the computer, the better the results get. Anderson considers the team's efforts akin to those of a painter who copies the masters to understand the artistic techniques used.
A brain can understand a word or sentence even if a signal or sound is partially obliterated, because the brain doesn't have just one way of understanding what's going on. Getting a traditional speech recognition program to handle multiple features and methods of analysis is difficult—if not impossible—with today's software and hardware.
Still, the goal remains to figure out how to teach a computer to analyze audio in the same way a human brain does. It could take thousands or millions of hours to train a system with enough samples for that kind of recognition. Anderson hopes that techniques similar to those used in enhancing computer images will help, providing a signal boost based upon small, recognized parts of a signal with output weighted and combined to make what the computer hears far more accurate.
While some simple applications of the machine hearing chip are already up and running, the more far-out applications—such as the Star Trek–esque computer voice recognition we all crave—are still years away. The machine needs to hear us before it can understand.
