6 The Cocktail Party Problem

At a crowded cocktail party, it can be a challenge to hear the person in front of you when the air is filled with a cacophony of others talking around you. Having two ears helps direct your hearing in the right direction, and your memory can fill in missing snatches of conversation. Now imagine a cocktail party with 100 people in a room and 100 nondirectional microphones spread around, each picking up sounds from everyone but with different ratios of amplitudes for each person on each microphone. Is it possible to devise an algorithm that can separate each of the voices into separate output channels? To make it even more difficult, what if the sound sources were unknown—such as music, clapping, nature sounds, or even random noise? This is called the “blind source separation problem” (figure 6.1).