New Scientist report (2000)

Homing in on a voice in the crowd

11 November 2000

Joanna Marchant

A hearing aid that can pick out a single voice from many different conversations-as easily as tuning a radio-could result from a new signal-processing technique.

Although people can easily pick out individual voices in a crowded room, repeating the feat artificially is tricky. The traditional method, called independent component analysis, works by assuming that different sources generate signals that change independently. For example, if two people speak at the same time, one voice does not depend on the other.

ICA is powerful but statistically complex, so computers using it take a long time to reach a conclusion, says psychologist Jim Stone of the University of Sheffield. Stone's method draws on how the real world works. "The physical world varies smoothly, so that's what inputs should do too" he says. "Most physical quantities, such as a single voice, tend to vary smoothly from second to second, while the babble of several people speaking at once is more complex." His technique teases out single voices by looking for signals that change smoothly over time, avoiding the sudden jumps that result when separate signals combine.

Because Stone's technique looks for a simple aspect of signals, software using it works very fast. He says that brain-imaging systems could also use it to pick out crucial signals from many others. It could even speed up the Internet by letting more signals travel at once down a wire or a fibre.

From New Scientist magazine, 168(2264), 16.

© Copyright New Scientist, RBI Ltd 2000