Dithering scientists make a sound profit from loss

Helena Flusfeder reports on compression that uses noise to improve quality

Researchers at Tel Aviv University have found that noise can improve music. But it has to be the right kind of noise.

Digital information is often compressed to squeeze it onto a disc or through a low-capacity channel such as a telephone line. "Lossless" compression techniques are used for text, computer programs and anything that must be reconstructed exactly. Sounds and images, however, are often compressed more aggressively by "lossy" techniques which involve a loss of quality.

By deliberately introducing computer-generated noise before the compression process, Meir Feder and his team have managed to reduce the distortions introduced when CD-quality music is compressed.

"We can control the way the noise distortion will come out," Professor Feder said. With PhD student Ram Zamir and several other research students he developed the idea of adding a small noise or disturbance called "dither" to sound before compressing it. The pay-off can be taken in higher sound quality or greater compression. The researchers say they could fit all Beethoven's symphonies on one CD. The method was originally introduced for image compression at the Massachusetts Institute of Technology in the 1960s. Professor Feder and his staff have adapted it for sound. He attributes his success to "the way of controlling the spectrum of the dither; of shaping the spectrum so that it fits better the way people hear."

Dither helps to overcome the "contour" or "stair" effects which occur when infinitely varying image colours and sound waveforms are quantised or approximated by a limited number of digital values.

Professor Feder explains: "Regular quantisation creates contour effects when applied to images and metallic effects when applied to sound - but if you add random noise on purpose, you get rid of these effects. What is important in signal compression is not just the amount of the distortion, but also the quality of the distortion. In our method, even if we keep some amount of the distortion or error, the quality of the signal is better."

Haim Bechor, manager of the central lab of Israel's government-supported multimedia consortium MOST, stressed the importance to the multimedia industry of being able to compress data and still get good quality: "Sound compression can reduce the amount of material going through and allow multimedia to be used before we have fibre optics."

The Internet is changing the requirements for data compression. "Whereas with telephone lines, the assumption by researchers was that the bit rate was constant, on the Internet the bit rate is constantly changing," Bechor said. "It is more of a random process: the bit rate goes up and down. The main message of companies like VocalTec and VDOnet is that they mastered the problem of the changing bit rate."

Professor Feder's research team has received grants from Intel, Inmarsat and Motorola which were spent on new and re-equipped laboratories. When Intel developed its MMX (multimedia extensions) processor, the company asked Professor Feder to help it exploit the power of the new processor for signal compression.

The Tel Aviv researchers developed algorithms which enable an MMX processor to perform four or five times faster than a Pentium. The result is that a PC can now decompress and display video in real time without special hardware.