Share on Facebook Tweet on Twitter Share on LinkedIn Share by email
Defeating Ambient Noise
By Suzanne Ross
July 19, 2004 12:00 AM PT

In the 1950s, Erector sets were the geekiest toy around. The hardest thing to build with an Erector set was a robot with a real motor. You were super cool if you could build the robot.

Ivan Tashev, Rico Malvar and his team have one-upped the robot. They built a microphone array from an Erector set, added some sophisticated software algorithms, and designed a microphone array that produces clear sound, free of background noise.

"It's a great way to build funny looking structures that don't exist yet," said Malvar.

Of course, that was just the prototype. Since the first Erector set version, they've refined the microphone array into several designs - from a sleek rectangular version that sits on your desk, to a built-in version for your laptop, or an amazing disc-shaped one for meeting rooms.

The goal of the Collaboration, Communication, and Signal Processing group was to build a microphone array that equaled the sound quality of a close-talk microphone, but allowed the user to move around the room without being attached to any wires.

Tashev thought that an array of microphones, combined with some software, could improve the signal to noise ratio. The signal to noise ratio, measured in decibels (dB) is the ratio of electronic noise that is present to some extent in all audio signals.

"The first microphone arrays were developed for submarines. The problem was that they required heavy computation," said Tashev. More recently, other companies have built microphone arrays that work well, but use a built-in processor and a power supply, which makes the unit very expensive.

"Microphones don't provide enough quality. What is the enemy? Noise. The ambient noise that surrounds us. We have two ears with 100 billion neurons between them that can do the job to filter out and reduce the noises around us. This is not the case for the single microphone," said Tashev.

The solution is to use a microphone array, which is a set of microphones positioned together. "By combining the signals from each microphone, you can make the sound quality almost equal to the shotgun like microphones in 'spy' movies that can work from 100 yards away," said Tashev.

The software that Tashev developed for the microphone array is able to pinpoint where the sound is coming from and amplify the voice while reducing ambient noise such as air conditioners and computer fans. The microphone array processing system reduces ambient noises between 12 and 30 dB. Right now, most computer microphones can only reduce the noise by about 6 to 10 dB.

The results from the four element array were better than an expensive commercial eight element array that contained its own processor. The only microphone that outperformed this microphone array was a close-up microphone.

Taking the Sound Out
The microphone array, combined with algorithms developed by the CCSP group, can do something that seems like magic to the average person. It can remove sound. For instance, if you were playing some music in your office, and you received a Windows messenger call, you would usually turn off the music. But with the acoustic echo cancellation algorithms in the microphone array, you could leave the music on. The acoustic echo canceller would reduce the background music between 12 and 30 dB.

"The echo cancellation can sound like magic," said Malvar. "It's useful, but it's not glitzy. On the other hand, it's real and it works and it will actually help people."

Recognizing Speech
The group also wanted their microphone array to advance speech recognition for PCs. One of the biggest challenges to developing good speech recognition was the input - if it wasn't clear the algorithms couldn't translate the sounds accurately.

The improved sound quality of the microphone arrays will help to improve speech recognition, though the researchers are still working towards their goal.

"We're down to about a 13 or 14 % word error rate, but the close-talk microphone has a 6% error rate. We hope that at the end of this summer we will have the same quality as the close-talk microphone. We're going to take the next serious step in the improvement of this technology," said Tashev.

"Signal processing has been around for 50 years, it's just beginning to mature. The innovation in signal processing is about how we use the technology to improve our day-to-day working on the computer," concluded Malvar.